Abstract

摘要

Primary bone tumors (PBTs) present significant diagnostic challenges due to their heterogeneous nature and similarities with bone infections. This study aimed to develop an ensemble deep learning framework that integrates multicenter radiographs and extensive clinical features to accurately differentiate between PBTs and bone infections.

原发性骨肿瘤（PBTs）因其异质性及与骨感染的相似性而带来了显著的诊断挑战。本研究旨在开发一种集成深度学习框架，整合多中心X光片和广泛的临床特征，以准确区分PBTs和骨感染。

We compared the performance of the ensemble model with four imaging models based solely on radiographs utilizing EfficientNet B3, EfficientNet B4, Vision Transformer, and Swin Transformers. The patients were split into external dataset (.

我们比较了集成模型与四个仅基于射线照片的影像模型的性能，这些模型利用了EfficientNet B3、EfficientNet B4、Vision Transformer和Swin Transformers。患者被分为外部数据集（。

N

北

= 423) and internal dataset [including training (

= 423) 以及内部数据集 [包括训练 (

N

北

= 1044), test (

= 1044)，测试 (

N

北

= 354), and validation set (

= 354)，以及验证集 (

N

北

= 171)]. The ensemble model outperformed imaging models, achieving areas under the curve (AUCs) of 0.948 and 0.963 on internal and external sets, respectively, with accuracies of 0.881 and 0.895. Its performance surpassed junior and mid-level radiologists and was comparable to senior radiologists (accuracy: 83.6%).

集合模型的表现优于影像模型，在内部和外部数据集上曲线下的面积（AUC）分别达到0.948和0.963，准确率分别为0.881和0.895。其性能超过了初级和中级放射科医生，并且与高级放射科医生相当（准确率：83.6%）。

These findings underscore the potential of deep learning in enhancing diagnostic precision for PBTs and bone infections (Research Registration Unique Identifying Number (UIN): researchregistry10483 and with details are available at .

这些发现强调了深度学习在提高 PBT 和骨感染诊断准确性方面的潜力（研究注册唯一识别号 (UIN)：researchregistry10483，详细信息可在 。

https://www.researchregistry.com/register-now#home/registrationdetails/6693845995ba110026aeb754/

https://www.researchregistry.com/register-now#home/registrationdetails/6693845995ba110026aeb754/

).

）。

Introduction

简介

Primary bone tumors (PBTs) are a diverse group of heterogeneous tumors that primarily develop in the skeletal system

原发性骨肿瘤（PBTs）是一组多样化的异质性肿瘤，主要发生在骨骼系统中。

1

1

. Despite their relatively low incidence, these malignancies present significant morbidity and mortality rates

尽管这些恶性肿瘤的发病率相对较低，但它们的发病和死亡率却很高。

2

2

,

，

3

3

. Remarkably, bone tumors rank as the third leading cause of cancer-related deaths among individuals under the age of 20 in the United States

值得注意的是，在美国，骨肿瘤在20岁以下人群中是癌症相关死亡的第三大原因。

4

4

. Currently, the treatment options for bone tumors remain formidable, traditional treatment options such as chemotherapy and surgical interventions, face significant challenges

目前，骨肿瘤的治疗选择仍然严峻，传统的治疗选择如化疗和手术干预面临着重大的挑战。

1

1

,

，

5

5

. For instance, chemotherapy often leads to severe side effects and has a limited success rate due to chemoresistance in specific type of bone tumors like osteosarcoma

例如，化疗常常导致严重的副作用，并且由于特定类型的骨肿瘤（如骨肉瘤）中的化学抗性，其成功率有限。

6

6

,

，

7

7

, while surgical options may result in functional impairments, residual metastasis, and even deformities or disabilities

，而手术选项可能导致功能障碍、残留转移，甚至畸形或残疾。

8

8

,

，

9

9

,

，

10

10

. These challenges underscore the need for improved treatment strategies. Radiography is the suggested primary auxiliary examination choice and commonly employed in orthopedic diagnosis as they generally provide a clear evaluation of the lesion’s location, internal matrix, margins, and associated periosteal reactions.

这些挑战强调了改进治疗策略的必要性。放射摄影是建议的主要辅助检查选择，在骨科诊断中普遍使用，因为它们通常能清晰评估病灶的位置、内部结构、边缘及相关的骨膜反应。

11

11

. These destruction signs reflect the biological activity of the lesion, thus allowing for evaluation of the malignancy assessment

这些破坏迹象反映了病灶的生物活性，从而允许对恶性评估进行评价。

12

12

. However, PBTs exhibit diverse compositions and may present with overlapping radiological and histological features

然而，PBTs表现出多样的组成，并可能呈现重叠的影像学和组织学特征。

13

13

,

，

14

14

. Consequently, the same PBTs may appear differently on radiographs, and different PBTs may exhibit similar radiographic images

因此，相同的PBT在射线照片上可能表现不同，而不同的PBT可能表现出相似的射线图像。

15

15

. Due to the rarity of PBTs, cultivating a professional radiologist often encounters the problem of a long training cycle and insufficient expertise

由于PBT较为罕见，培养专业的放射科医生常遇到培训周期长、专业知识不足的问题。

16

16

. Bone infections primarily encompass osteomyelitis and joint infections. Notably, clinically distinguishing PBTs from bone infections is challenging for the similarities in clinical practice (e.g., fever, soft tissue swelling, periosteal reaction), leading to potential confusion and challenges in accurate diagnosis.

骨感染主要包括骨髓炎和关节感染。值得注意的是，由于临床实践中的相似性（如发热、软组织肿胀、骨膜反应），将PBTs与骨感染区分开来具有挑战性，这可能导致潜在的混淆，并给准确诊断带来困难。

17

17

,

，

18

18

. Therefore, the preoperative differential diagnosis of PBTs and bone infections is crucial for precise diagnosis and timely treatment.

因此，PBTs和骨感染的术前鉴别诊断对于精确诊断和及时治疗至关重要。

Traditional diagnostic methods heavily rely on the expertise and subjective judgment of radiologists and pathologists, which can lead to potential errors and delays in treatment options

传统诊断方法严重依赖放射科医生和病理学家的专业知识和主观判断，这可能导致潜在的错误和治疗选择的延误。

19

19

,

，

20

20

,

，

21

21

. Furthermore, if imaging studies are not interpreted by musculoskeletal radiologists who specialize in this field, discrepancies in readings can occur, reaching up to 28%

此外，如果影像学检查不是由专门从事此领域的肌肉骨骼放射科医生解读，可能会出现高达28%的读数差异。

22

22

. In recent years, the emergence of deep learning algorithms especially convolutional neural networks (CNNs) has significantly impacted clinical practices such as assisted diagnosis and drug discovery

近年来，深度学习算法特别是卷积神经网络（CNNs）的出现显著影响了临床实践，如辅助诊断和药物发现等领域。

23

23

,

，

24

24

. These advancements have also demonstrated improvements in cancer prognosis

这些进展还展示了癌症预后的改善

25

25

. The application of deep learning in cancer diagnosis has considerably enriched the field, showcasing astounding efficiency in solving complex problems with a lower error rate than humans

深度学习在癌症诊断中的应用极大地丰富了这一领域，以比人类更低的错误率解决了复杂问题，展现了惊人的效率。

26

26

,

，

27

27

. For bone tumors, the development of multitask deep learning models has enabled accurate and simultaneous bounding box placement and segmentation of PBTs in radiographs, and can effectively differentiate benign and malignant PBTs with performance comparable to senior radiologists

对于骨肿瘤，多任务深度学习模型的发展已经能够实现对放射影像中PBT的精准同步边界框定位和分割，并能有效区分良性和恶性PBT，其性能可与资深放射科医生相媲美。

28

28

. Due to the rarity of PBTs, deep learning models in this domain are constrained by limited access to large-scale cohort datasets, resulting in scant efforts to differentiate between bone tumors and other bone pathologies. Furthermore, prevailing models emphasize algorithmic versatility and data diversity, yet they fall short in sufficiently incorporating crucial clinical patient data and prioritizing the interpretability of model outcomes.

由于PBTs罕见，该领域的深度学习模型受限于无法获取大规模队列数据集，导致区分骨肿瘤与其他骨病的努力甚少。此外，现有模型强调算法的通用性和数据多样性，但在充分整合关键临床患者数据以及优先考虑模型结果的可解释性方面存在不足。

This trend runs counter to the fundamental ethos of algorithmic design, sometimes it is necessary to pause and delve into a profound comprehension of our meticulously crafted models with professional radiologist interpretation, thereby aligning our efforts with the original essence of algorithmic innovation..

这一趋势与算法设计的基本理念背道而驰，有时我们需要暂停下来，通过专业放射科医生的解读深入理解我们精心构建的模型，从而将我们的努力与算法创新的初衷重新对齐。

Therefore, the main objective of this study was to create an ensemble deep learning framework using multicenter radiographs and extensive clinical features to accurately differentiate between PBTs and bone infections. While comparing the performance of the ensemble model with four imaging models merely utilizing radiographs, which were built upon four distinct neural networks: EfficientNet B3 (E3), EfficientNet B4 (E4), Vision Transformer (ViT), and Swin Transformers (SWIN).

因此，本研究的主要目标是创建一个集成深度学习框架，利用多中心射线照片和广泛的临床特征，以准确区分PBTs和骨感染。在比较集成模型与四个仅使用射线照片的成像模型的性能时，这四个模型基于四种不同的神经网络构建：EfficientNet B3（E3）、EfficientNet B4（E4）、Vision Transformer（ViT）和Swin Transformers（SWIN）。

Subsequently, these models’ effectiveness was assessed and compared with the diagnostic accuracy of radiologists. In addition, six professional radiologists, categorized into three seniority groups, provided insights and discussions on the clinical implications of the developed models. The research methodology and study flowchart are illustrated in Fig.

随后，评估了这些模型的有效性，并将其与放射科医生的诊断准确性进行了比较。此外，六位专业放射科医生被分为三个不同资历组，对所开发模型的临床意义提供了见解和讨论。研究方法和流程图如图所示。

.

。

1

1

.

。

Fig. 1: Design and flowchart of the deep learning framework.

图1：深度学习框架的设计和流程图。

a

a

Preprocessing of data. The input of the models mainly includes image information based on radiographs defined as input (A) and clinical information defined as input (B).

数据预处理。模型的输入主要包括基于射线照片的图像信息，定义为输入（A），以及临床信息，定义为输入（B）。

b

b

Model development.

模型开发。

c

c

Comprehensive prediction. P

全面预测。P

Radio

收音机

and P

和 P

Clinic

诊所

refers to the results of the four imaging models (E3, E4, ViT, and SWIN) and the clinic model, respectively.

分别指四个影像模型（E3、E4、ViT 和 SWIN）和临床模型的结果。

d

d

Evaluation. This part is mainly composed of ROC curve and confusion matrix.

评估。这一部分主要由ROC曲线和混淆矩阵组成。

e

e

Verifying. The results of models are compared with radiologists with different seniority.

验证。将模型的结果与不同资历的放射科医生进行比较。

n

n

number of the radiographs, E3 EfficientNet B3, E4 EfficientNet B4, ViT vision transformer, SWIN swin transformers. Note: Fig. 1 was Created with BioRender.com.

放射图像的数量，E3 EfficientNet B3，E4 EfficientNet B4，ViT视觉Transformer，SWIN Swin Transformer。注：图1由BioRender.com创建。

Full size image

全尺寸图像

Results

结果

Characteristics of study participants

研究参与者的特点

This retrospective study included 1992 patients (median age, 29 years; range, 1–88 years; 796 female) from three hospitals diagnosed of PBTs or bone infections with histopathology reports available as reference (Table

这项回顾性研究包括来自三家医院的1992名患者（中位年龄，29岁；范围，1-88岁；796名女性），这些患者被诊断为PBTs或骨感染，并有组织病理学报告作为参考（表

1

1

). The distribution of 1208 patients with PBTs were described in Supplementary Table

). 1208名PBTs患者的分布情况描述于补充表格中。

1

1

, with 767 benign subtypes, 251 malignant subtypes and 190 intermediate subtypes according to the 2020 World Health Organization (WHO) system for the classification for tumors of bone. While for 784 patients with bone infection, bone tuberculosis counted the highest proportion (Supplementary Table

，根据2020年世界卫生组织（WHO）骨肿瘤分类系统，包含767种良性亚型、251种恶性亚型和190种中间型亚型。而在784例骨感染患者中，骨结核占比最高（补充表）。

2

2

). 1569 patients from Hospital 1 were utilized as internal dataset and divided into a training set (

). 医院1的1569名患者被用作内部数据集，并分为训练集 (

N

北

= 1044), a test set (

= 1044)，测试集（

N

北

= 354) and a validation set (

= 354) 以及一个验证集 (

N

北

= 171) (Fig.

= 171) （图。

2a

2a

) (screening criteria in Fig.

)（图中的筛选标准

2b

2b

); 423 patients from Hospital 2 and Hospital 3 were used for external validation (Supplementary Fig.

); 来自医院2和医院3的423名患者用于外部验证（补充图。

1

1

). Clinical characteristics like age, lesion location, pain, swelling, trauma, C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), alkaline phosphatase (ALP) among all of the bone infection and PBT patients had significantly different distributions (Table

）。所有骨感染和PBT患者在年龄、病灶位置、疼痛、肿胀、创伤、C反应蛋白（CRP）、红细胞沉降率（ESR）、碱性磷酸酶（ALP）等临床特征上均有显著不同的分布（表

1

1

). The clinical characteristics of patients with PBTs and bone infection were summarized specifically in Supplementary Tables

）。PBTs和骨感染患者的临床特征具体总结在补充表格中。

3

3

and

和

4

4

. We further found that clinical characteristics like age, lesion location, pain, swelling, trauma, C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), alkaline phosphatase (ALP) also had statistical differences in the internal dataset (Supplementary Table

我们进一步发现，年龄、病灶位置、疼痛、肿胀、创伤、C反应蛋白（CRP）、红细胞沉降率（ESR）、碱性磷酸酶（ALP）等临床特征在内部数据集中也存在统计学差异（补充表）。

5

5

).

）。

Table 1 Clinical characteristics of included patients with primary bone tumors or bone infections

表1 原发性骨肿瘤或骨感染患者的临床特征

Full size table

全尺寸表格

Fig. 2: Data distribution and the screening criteria of the study.

图 2：研究的数据分布和筛选标准。

a

a

Data processing process and data distribution across different datasets.

数据处理流程以及跨不同数据集的数据分布。

b

b

Screening criteria of the research.

研究的筛选标准。

n

n

number of the radiographs,

放射图像的数量，

N

北

number of the patients. Note: Fig. 2 was Created with BioRender.com.

患者数量。注：图2由BioRender.com创建。

Full size image

全尺寸图像

Classification performance of models

模型的分类性能

In the internal test set, the ensemble model outperformed four imaging models (E3, E4, ViT and SWIN) on the binary classification to distinguish PBTs from bone infections respectively (

在内部测试集中，集成模型在区分PBTs和骨感染的二元分类上分别优于四种影像学模型（E3、E4、ViT和SWIN）(

P

P

< 0.001 for E3, E4, and ViT;

E3、E4 和 ViT 的值均 < 0.001；

P

P

= 0.835 for SWIN; DeLong test) (Table

= 0.835（针对SWIN；DeLong检验）（表

2

2

and Supplementary Fig.

和补充图。

2

2

). Specifically, the ensemble model reached an AUC of 0.948 (95% CI, 0.931–0.963) and an accuracy of 88.1% for binary classification, whereas the E3, E4, ViT and SWIN-based models achieved AUCs of 0.903 (95% CI, 0.878–0.927), 0.912 (95% CI, 0.890–0.934), 0.903 (95% CI, 0.880–0.927), and 0.946 (95% CI, 0.929–0.963) as well as accuracies of 84.3%, 84.6%, 84.3%, and 87.2%, respectively (Table .

）。具体来说，集成模型在二元分类中达到了0.948的AUC（95%置信区间，0.931–0.963）和88.1%的准确率，而基于E3、E4、ViT和SWIN的模型分别达到了0.903（95%置信区间，0.878–0.927）、0.912（95%置信区间，0.890–0.934）、0.903（95%置信区间，0.880–0.927）和0.946（95%置信区间，0.929–0.963）的AUC，以及84.3%、84.6%、84.3%和87.2%的准确率（表 。

2

2

). The ROC curves and the confusion matrices also demonstrated the best categorizing ability of the ensemble model (Fig.

）。ROC曲线和混淆矩阵也展示了集成模型的最佳分类能力（图。

3

3

and Supplementary Fig.

和补充图。

3

3

).

）。

Table 2 Performance of the models and radiologists of different seniority in internal and external test set

表2 不同资历的模型和放射科医生在内部和外部测试集中的表现

Full size table

全尺寸表格

Fig. 3: Confusion matrix and receiver operating characteristic (ROC) curve of the ensemble model for the binary classification.

图3：二元分类的集成模型的混淆矩阵和接收者操作特征（ROC）曲线。

a

a

,

，

b

b

ROC curve and confusion matrices of all models and radiologists’ interpretations on the internal test set.

所有模型和放射科医生在内部测试集上的ROC曲线和混淆矩阵。

c

c

,

，

d

d

ROC curve and confusion matrices of all models on the external test set. Note: EG1= expert 1+ expert 2 (junior radiologist group); EG2= expert 3+ expert 4 (medium seniority group); EG3= expert 5+ expert 6 (senior radiologist group). EG expert group, E3 EfficientNet B3, E4 EfficientNet B4, ViT vision transformer, SWIN swin transformers, AUC area under the curve, Acc accuracy..

所有模型在外部测试集上的ROC曲线和混淆矩阵。注：EG1=专家1+专家2（初级放射科医生组）；EG2=专家3+专家4（中高级别组）；EG3=专家5+专家6（高级放射科医生组）。EG 专家组，E3 EfficientNet B3，E4 EfficientNet B4，ViT 视觉Transformer，SWIN Swin Transformer，AUC 曲线下面积，Acc 准确率。

Full size image

全尺寸图像

In the external test set for validation, the ensemble model also outperformed the four imaging models, which proved the consistency and applicability of the ensemble model (

在用于验证的外部测试集中，集成模型同样优于四个成像模型，这证明了集成模型的一致性和适用性 (

P

P

< 0.001 for E3 and E4;

E3 和 E4 的值均小于 0.001；

P

P

= 0.002 for ViT and SWIN; DeLong test) (Table

= 0.002 对于 ViT 和 SWIN；DeLong 检验）（表

2

2

and Supplementary Fig.

和补充图。

2

2

). Specifically, the ensemble model reached an AUC of 0.963 (95% CI, 0.951–0.973) and an accuracy of 89.5% for the classification, while the four imaging models reached AUCs of 0.930 (95% CI, 0.914–0.946), 0.946 (95% CI, 0.932–0.960), 0.951 (95% CI, 0.939–0.964), and 0.957 (95% CI, 0.944–0.969) as well as accuracies of 86.6%, 87.4%, 87.1%, and 88.5%, respectively (Table .

）。具体而言，集成模型在分类任务中达到了0.963的AUC（95%置信区间，0.951–0.973）和89.5%的准确率，而四个影像模型分别达到了0.930（95%置信区间，0.914–0.946）、0.946（95%置信区间，0.932–0.960）、0.951（95%置信区间，0.939–0.964）和0.957（95%置信区间，0.944–0.969）的AUC，以及86.6%、87.4%、87.1%和88.5%的准确率（表 。

2

2

). The confusion matrices and ROC curves in Fig.

）。图中的混淆矩阵和ROC曲线

3

3

further visually demonstrated the superior discrimination capability of the ensemble framework. In addition, the result in internal validation set further confirmed the stability and consistency of the ensemble model (Supplementary Fig.

进一步直观地展示了集成框架的优越区分能力。此外，内部验证集的结果进一步证实了集成模型的稳定性和一致性（补充图）。

4

4

).

）。

Comparison of performance between the ensemble framework and radiologists

集成框架与放射科医生的性能比较

In this study, six professional radiologists were divided into junior expert group (EG1), medium seniority group (EG2), and senior expert group (EG3). The comparative analysis was conducted using the internal test set. As shown in Fig.

在本研究中，六名专业放射科医生被分为初级专家组（EG1）、中级资深组（EG2）和高级专家组（EG3）。对比分析使用内部测试集进行。如图所示。

3

3

, the ensemble framework significantly outperformed all three radiologist groups (

，集成框架显著优于所有三组放射科医生（

P

P

< 0.001 for EG1, EG2, and EG3; Cochran’s Q test) (Table

<0.001 for EG1, EG2, and EG3; Cochran’s Q test) （表

2

2

). The SWIN-based imaging model demonstrated comparable performance to the ensemble model (

基于SWIN的成像模型表现出与集成模型相当的性能（

P

P

= 0.835; DeLong test) (Table

= 0.835；DeLong 检验）（表

2

2

) and also outperformed the three radiologist groups. The other three imaging models (E3, E4, and ViT) achieved superior performance compared to EG1 and EG2, and were comparable to EG3. In addition, we calculated and provided other metrics, including accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 Score, to facilitate a comprehensive comparison of the performance between the ensemble framework and the radiologists (Table .

）并且还超过了三组放射科医生的表现。其他三个成像模型（E3、E4 和 ViT）相较于EG1和EG2表现出更优的性能，并且与EG3相当。此外，我们计算并提供了其他指标，包括准确性、敏感性、特异性、阳性预测值（PPV）、阴性预测值（NPV）和F1分数，以便于对集成框架和放射科医生之间的表现进行全面比较（表 。

3

3

).

）。

Table 3 Performance of the experts and models in classifying high-frequency lesions in PBTs and bone infections in the internal test set

表3 专家和模型在内部测试集中对PBT高频病变和骨感染分类的性能

Full size table

全尺寸表格

Inter-reader reliability

读者间可靠性

Considering the subjectivity of individual sample predictions and large workload of the monotonous radiographs (

考虑到单个样本预测的主观性和单调的射线照片的大量工作负载 (

n

n

= 687), inter-reader reliability among radiologists was much lower than that of the models. We compared the best performing model—the ensemble model with experts of diverse seniority, Cohen

= 687)，放射科医生之间的阅片者间可靠性远低于模型。我们比较了表现最佳的模型——由不同资历专家组成的集成模型，Cohen

κ

κ

between expert 6 (radiologist with the highest seniority) and the ensemble had the best consistency: 0.596 (95% CI, 0.560–0.633) (Table

专家6（资历最高的放射科医生）与整体评估之间的一致性最佳：0.596（95% CI，0.560–0.633）（表

4

4

). The Fleiss

). 弗莱斯

κ

κ

value among radiologists achieved 0.401 (95% CI, 0.364–0.438) on the internal test set, while the Fleiss

放射科医生之间的价值在内部测试集上达到了0.401（95%置信区间，0.364-0.438），而Fleiss

κ

κ

value among models achieved 0.800 (95% CI, 0.770–0.830) (Table

模型之间的价值达到了0.800（95%置信区间，0.770-0.830）（表

4

4

). Furthermore, we used Cohen

）。此外，我们使用了Cohen

κ

κ

value to evaluate consistency between pairs of expert groups (EG1, EG2, and EG3) and consistency between the ensemble model and the other four imaging models. We found as seniority increased, the consistency of judgment rose in radiologists, but the overall consistency of judgment was still lower than that of the models.

评估三组专家（EG1、EG2 和 EG3）之间的一致性以及集成模型与另外四个影像模型之间的一致性。我们发现，随着资历的增加，放射科医生的判断一致性提高，但总体判断一致性仍低于模型。

The Fleiss .

弗莱斯。

κ

κ

value among EG1, EG2, and EG3 reached 0.267 (95% CI, 0.234–0.300), 0.295 (95% CI, 0.261–0.329), and 0.581 (95% CI, 0.544–0.618), respectively (Table

EG1、EG2 和 EG3 的值分别达到 0.267（95% CI，0.234–0.300）、0.295（95% CI，0.261–0.329）和 0.581（95% CI，0.544–0.618）（表

4

4

). In contrast, the Fleiss κ value among the ensemble model and the imaging models reached 0.805 (95% CI, 0.775–0.835), 0.793 (95% CI, 0.763–0.823), 0.783 (95% CI, 0.752–0.814), and 0.908 (95% CI, 0.886–0.930), respectively (Table

）。相比之下，集成模型与影像模型之间的Fleiss κ值分别达到了0.805（95% CI，0.775–0.835）、0.793（95% CI，0.763–0.823）、0.783（95% CI，0.752–0.814）和0.908（95% CI，0.886–0.930）（表

4

4

). This indicates that a strong disagreement exists among junior radiologists when facing classification of PBTs and bone infection solely on radiograph data.

). 这表明，初级放射科医生在仅根据X光片数据面对PBTs和骨感染分类时存在强烈分歧。

Table 4 Inter-reader reliability of the models and radiologists

表4 模型与放射科医生的阅片者间可靠性

Full size table

全尺寸表格

Visual interpretation of models

模型的可视化解释

In order to accurately interpret the predictions made by the models, we employed techniques such as GradCAM and ScoreCAM to visualize the specific regions within the input data that the model utilizes for its decision-making process (Fig.

为了准确解释模型的预测结果，我们采用了 GradCAM 和 ScoreCAM 等技术来可视化输入数据中模型用于决策过程的具体区域（图。

4

4

). By identifying and highlighting these key areas, we are able to gain a deeper understanding of how the model arrives at its predictions and make informed assessments about its performance and reliability. In general, the analysis of the highlighted regions on the heat maps reveals that the model primarily focused on identifying PBT or bone infection lesions, such as hemorrhage, necrosis, calcification, cystic lesions, and inflammatory exudation.

）。通过识别和突出这些关键区域，我们能够更深入地了解模型如何得出其预测，并对其性能和可靠性做出明智的评估。总体而言，对热图上突出显示区域的分析表明，该模型主要专注于识别PBT或骨感染病灶，例如出血、坏死、钙化、囊性病变和炎性渗出。

These findings are in line with the segmentation results, indicating that the model was able to achieve a high level of accuracy in classifying these specific types of lesions. This demonstrates the effectiveness of the model in accurately identifying and categorizing pathological features, ultimately leading to satisfactory classification performance.

这些发现与分割结果一致，表明模型能够准确地对这些特定类型的病灶进行分类。这证明了该模型在准确识别和分类病理特征方面的有效性，最终取得了令人满意的分类性能。

The distinctions between GradCAM and ScoreCAM are clearly evident in the generated heat maps. GradCAM primarily emphasizes the areas of bone hyperplasia and sclerosis, neglecting those of bone destruction. Conversely, ScoreCAM directs its attention toward both osteogenic and osteoclastogenic regions, resulting in a more precise delineation of lesion boundaries..

GradCAM 和 ScoreCAM 之间的区别在生成的热力图中表现得非常明显。GradCAM 主要强调骨增生和硬化区域，而忽略了骨破坏区域。相反，ScoreCAM 则同时关注成骨和破骨区域，从而更精确地描绘了病灶边界。

Fig. 4: Visualization of PBTs and bone infections in four cases respectively.

图4：分别显示了四个病例的PBT和骨感染情况。

a

a

Visualization of PBTs. Patient 1, a 10-year-old girl with chondrosarcoma on the left proximal humerus; Patient 2, a 10-year-old boy with a simple bone cyst on the right humerus; Patient 3, a 65-year-old female with giant cell tumor of bone on the left distal femur; Patient 4, a 9-year-old boy with osteosarcoma on the left distal femur.

PBT的可视化。患者1，一名10岁女孩，左肱骨近端患有软骨肉瘤；患者2，一名10岁男孩，右肱骨患有单纯性骨囊肿；患者3，一名65岁女性，左股骨远端患有骨巨细胞瘤；患者4，一名9岁男孩，左股骨远端患有骨肉瘤。

.

。

b

b

Visualization of Bone infection. Patient 5, a 72-year-old male with chronic suppurative osteomyelitis of the lower right femur; Patient 6, a 31-year-old male with tuberculosis of lumbar vertebrae 3 and 4 with spinal canal stenosis; Patient 7, a 68-year-old female with tuberculosis of left knee joint; Patient 8, a 65-year-old male with right distal femoral osteomyelitis.

骨感染的可视化。患者5，72岁男性，右股骨下段慢性化脓性骨髓炎；患者6，31岁男性，第3、4腰椎结核伴椎管狭窄；患者7，68岁女性，左膝关节结核；患者8，65岁男性，右股骨远端骨髓炎。

Starting from the left, the first column is the original flat film image. The second column is an area cut as small as possible against the edge of the lesion. The third column is the GradCAM-generated heat map. The fourth is the heat map generated by ScoreCAM..

从左开始，第一列为原始平片图像。第二列为尽可能小地对病灶边缘进行裁剪的区域。第三列为GradCAM生成的热力图。第四列为ScoreCAM生成的热力图。

Full size image

全尺寸图像

Radiologist interpretation

放射科医生解读

Diagnosis of the ensemble model and radiologists across different types of PBTs and bone infections were explicated in Supplementary Tables

诊断集成模型和放射科医生在不同类型PBT和骨感染中的表现详见补充表格。

6

6

and

和

7

7

, specifically. Some bone tumors were classified incorrectly by experts but correctly by the model (Fig.

，具体来说。一些骨肿瘤被专家错误分类，但被模型正确分类（图。

5

5

). Giant cell tumors of bone (Fig.

). 骨巨细胞瘤（图。

5a

5a

) may exhibit obvious aggressiveness, resulting in the blurring of the boundary between the lesion and normal bone, wormlike and ethmoidal bone destruction, and soft tissue masses beyond the bone envelope. There is partial image overlap with malignant bone tumors and infections (such as Brodie abscess) on plain film.

）可能表现出明显的侵袭性，导致病灶与正常骨边界模糊、虫蚀样和筛孔样骨破坏以及超出骨包膜的软组织肿块。在平片上，与恶性骨肿瘤和感染（如Brodie脓肿）存在部分影像重叠。

29

29

. Synovial osteo-chondromatosis (Fig.

. 滑膜骨软骨瘤病（图。

5b

5b

) is characterized by multiple cartilage nodules in the joint lumen. When the cartilage nodules are not significantly calcified, especially when bone erosion is present at the same time, it is difficult to distinguish osteoarthritis with free bodies in the joint

）的特征是关节腔内存在多个软骨结节。当软骨结节没有显著钙化时，特别是同时存在骨侵蚀的情况下，很难区分骨关节炎和关节内游离体。

30

30

. There are also cases where both experts and models misclassify. Chondrosarcoma (Fig.

. 也有专家和模型都误分类的情况。软骨肉瘤（图。

5e

5e

) involving the pelvis is more likely to occur in the iliac wing than in the acetabulum. Intramedullary osteolytic lesions with poorly defined acetabular boundaries may be consistent with chondrosarcoma, as well as tuberculosis and osteoarthritis of the hip. The overlap of the structure in plain film makes the calcification of the circular or arc-shaped chondroid stroma, a typical manifestation of chondrosarcoma at the acetabulum, not obvious, and appears to be suspected involvement of the adjacent femoral head.

) 盆腔病变更易发生在髂骨翼而非髋臼。髓内溶骨性病变伴髋臼边界不清可能与软骨肉瘤、结核和髋关节骨关节炎相符。平片中结构的重叠使得环形或弧形软骨基质的钙化（软骨肉瘤在髋臼的典型表现）不明显，并且似乎怀疑累及相邻的股骨头。

Multiple myeloma (Fig. .

多发性骨髓瘤（图。

5f

5f

) tends to occur in the thoracic vertebrae and has a positive pedicle sign (destruction of the vertebral body but retention of the pedicle). When both the vertebral body and pedicle are destroyed at the same time, it is necessary to distinguish them from spinal metastasis and spinal tuberculosis with insignificant paravertebral abscess.

）易发生于胸椎，且椎弓根征阳性（椎体破坏但椎弓根保留）。当椎体与椎弓根同时破坏时，需要与脊柱转移瘤、椎旁脓肿不明显的脊柱结核相鉴别。

31

31

. There are also cases where the experts got the classification right and the model got it wrong. Sclerosing osteosarcoma has no obvious bone destruction, which is different from the common mixed osteosarcoma with both osteolytic and sclerosing (Fig.

. 还有一些案例是专家分类正确，而模型出错的情况。硬化性骨肉瘤没有明显的骨质破坏，这与常见的混合型骨肉瘤（兼具溶骨性和硬化性）不同（图。

5c

5c

). Giant cell tumors of bone occur mostly in the long bone, but can also occur in the vertebral body (Fig.

). 骨巨细胞瘤多发生于长骨，但也可能发生在椎体 (图。

5d

5天

). These relatively uncommon conditions can be recognized by radiologists with extensive clinical experience. However, due to limited training on rare cases, the model tends to focus more on interpreting the more frequently encountered chronic osteomyelitis and spinal tuberculosis.

）。这些相对少见的情况可以被具有丰富临床经验的放射科医生识别。然而，由于对罕见病例的训练有限，该模型更倾向于关注解释更为常见的慢性骨髓炎和脊柱结核。

Fig. 5: Bone tumor cases misclassified by experts and models in the internal test set.

图5：内部测试集中被专家和模型误分类的骨肿瘤病例。

a

a

,

，

b

b

The models mostly predict correctly but the experts mostly predict incorrectly based on the radiographs from Patient a and Patient b.

模型大多预测正确，但专家基于患者a和患者b的X光片大多预测错误。

c

c

,

，

d

d

The models mostly predict incorrectly but the experts mostly predict correctly based on the radiographs from Patient c and Patient d.

基于患者c和患者d的X光片，模型大多预测错误，但专家大多预测正确。

e

e

,

，

f

f

Both of the models and the experts mostly predict incorrectly based on the radiographs from Patient e and Patient f. Model classification shows the probability of SWIN model and E3 model, which respectively correspond to the best and worst predictions in the imaging models. Red circles refer to bone tumors.

基于患者e和患者f的X光片，两个模型和专家大多预测错误。模型分类显示了SWIN模型和E3模型的概率，它们分别对应影像模型中最佳和最差的预测。红色圆圈表示骨肿瘤。

Blue circles refer to bone infections. Bar = 100 μm. E3 EfficientNet B3, SWIN swin transformers, GCT giant cell of bone, SC Synovial chondromatosis, OS osteosarcoma, CS Chondrosarcoma, PC plasmacytoma. Note: Fig. 5 was Created with BioRender.com..

蓝色圆圈表示骨感染。标尺=100微米。E3 EfficientNet B3，SWIN Swin Transformers，GCT 骨巨细胞，SC 滑膜软骨瘤病，OS 骨肉瘤，CS 软骨肉瘤，PC 浆细胞瘤。注：图5使用BioRender.com创建。

Full size image

全尺寸图像

Some bone infections were classified incorrectly by experts but correctly by the model (Fig.

一些骨感染被专家错误分类，但被模型正确分类（图。

6

6

). Chronic sclerosing osteomyelitis (Garre osteomyelitis, Fig.

). 慢性硬化性骨髓炎（Garre骨髓炎，图。

6a

6a

) mainly presents with osteosclerosis and lack of dead bone formation, and needs to be distinguished from sclerosing osteosarcoma

）主要表现为骨硬化和无死骨形成，需要与硬化型骨肉瘤相鉴别。

32

32

. When lumbar tuberculosis (Fig.

. 当腰椎结核（图。

6b

6b

) involves only a single vertebral body and lacks paravertebral space narrowing, formation of paravertebral cold abscess, and soft tissue calcification, it should be differentiated from plasma-cell tumor and giant cell tumor of bone. There are also cases where the experts got the classification right, and the model got it wrong.

仅累及单个椎体，无椎旁间隙变窄、椎旁寒性脓肿形成和软组织钙化，应与浆细胞瘤和骨巨细胞瘤相鉴别。也有一些专家分类正确而模型分类错误的情况。

There is partial overlap between acute suppurative osteomyelitis (Fig. .

急性化脓性骨髓炎之间存在部分重叠（图。

6d

6天

) and Ewing sarcoma. Although the image manifestations of joint tuberculosis (Fig.

）和尤文肉瘤。尽管关节结核的影像学表现（图。

6c

6c

) occurring in the elbow joint are relatively typical, the number of training cases of joint tuberculosis in the extremities is limited for the model, and more common training cases of tuberculosis come from spinal tuberculosis, resulting in a decrease in the accuracy of model interpretation. There are also cases in which both experts and models misclassify.

）出现在肘关节相对典型，肢体关节结核的训练样本数量对于模型而言有限，更多常见的结核训练样本来自脊柱结核，导致模型判读的准确性下降。也有专家和模型均误判的情况。

Brodie abscess appears as a single osteolytic lesion on X-ray, accompanied by peripheral sclerosis with decreasing degree of peripheral sclerosis, which is difficult to distinguish from osteosarcoma and osteoid osteoma (Fig. .

Brodie脓肿在X光片上表现为单一的溶骨性病变，伴有逐渐减轻的周围硬化，这使得它很难与骨肉瘤和骨样骨瘤区分（图。

6f

6f

). When not accompanied by obvious sclerosis, it is difficult to distinguish Langerhans histiocytosis and Ewing sarcoma (Fig.

). 在没有明显硬化的情况下，很难区分朗格汉斯组织细胞增生症和尤文肉瘤（图。

6e

6e

)

)

33

33

.

。

Fig. 6: Bone infection cases misclassified by experts and models in the internal test set.

图 6：内部测试集中被专家和模型误分类的骨感染病例。

a

a

,

，

b

b

The models mostly predict correctly but the experts mostly predict incorrectly based on the radiographs from Patient g and Patient h.

根据患者g和患者h的X光片，模型大多预测正确，而专家大多预测错误。

c

c

,

，

d

d

The models mostly predict incorrectly but the experts mostly predict correctly based on the radiographs from Patient i and Patient j.

这些模型大多预测错误，但专家们根据患者i和患者j的X光片大多预测正确。

e

e

,

，

f

f

Both of the models and the experts mostly predict incorrectly based on the radiographs from Patient k and Patient l. Model classification shows the probability of SWIN model and E3 model, which respectively correspond to the best and worst predictions in the imaging models. Red circles refer to bone tumors.

基于患者k和患者l的X光片，两个模型和专家大多预测错误。模型分类显示了SWIN模型和E3模型的概率，这两个模型分别对应影像模型中最好和最差的预测结果。红色圆圈表示骨肿瘤。

Blue circles refer to bone infections. Bar = 100 μm. E3 EfficientNet B3, SWIN swin transformers, COM chronic osteomyelitis, LVT lumber vertebra tuberculosis, JT joint tuberculosis, OM osteomyelitis, BA brodie’s abscess. Note: Fig. 6 was Created with BioRender.com..

蓝圈表示骨感染。比例尺 = 100 μm。E3 EfficientNet B3，SWIN Swin Transformers，COM 慢性骨髓炎，LVT 腰椎结核，JT 关节结核，OM 骨髓炎，BA Brodie脓肿。注：图6使用BioRender.com创建。

Full size image

全尺寸图像

Discussion

讨论

Overall, our research introduced an innovative ensemble framework designed to detect and classify PBTs and bone infections concurrently. This framework incorporated two distinct single models: a radiograph-based imaging model and a clinical logistic regression model. By combining these models, we were able to enhance the classification accuracy of radiologists, surpassing the diagnostic capabilities of junior radiologists and aligning closely with those of medium senior radiologists.

总体而言，我们的研究引入了一种创新的集成框架，旨在同时检测和分类PBTs和骨感染。该框架包含了两个不同的单一模型：基于X光片的影像模型和临床逻辑回归模型。通过结合这些模型，我们能够提高放射科医生的分类准确性，超越初级放射科医生的诊断能力，并与中高级放射科医生的能力紧密匹配。

Our findings suggest that this ensemble approach holds promise for improving the accuracy and efficiency of detecting and classifying PBTs and bone infections in clinical settings..

我们的研究结果表明，这种集成方法有望提高在临床环境中检测和分类PBT和骨感染的准确性和效率。

In the realm of medical imaging, numerous deep learning models have been developed to aid in the diagnosis and classification of skeletal diseases using data from radiographs

在医学影像领域，已经开发了许多深度学习模型，以利用来自X光片的数据帮助诊断和分类骨骼疾病。

28

28

,

，

34

34

,

，

35

35

, CT

，CT

36

36

,

，

37

37

,

，

38

38

, and MRIs

，以及核磁共振成像

39

39

,

，

40

40

,

，

41

41

. However, the majority of these models have primarily concentrated on feature extraction from images and enhancing the accuracy of classification judgments to optimize model performance, neglecting the initial goal of utilizing deep learning as an auxiliary tool to enhance the diagnostic accuracy of clinicians.

然而，这些模型大多集中在图像特征提取和提高分类判断的准确性上，以优化模型性能，却忽视了利用深度学习作为辅助工具来提高临床医生诊断准确性的初衷。

Consequently, our study aims to shed light on this issue by employing GradCAM and ScoreCAM to visualize the areas of focus within the models. In the course of our research, we have observed that GradCAM tends to prioritize the identification of bone hyperplasia and sclerosis, while overlooking areas of bone destruction.

因此，我们的研究旨在通过使用GradCAM和ScoreCAM来可视化模型中的关注区域，以此阐明这一问题。在研究过程中，我们观察到GradCAM倾向于优先识别骨增生和骨硬化，而忽视了骨破坏区域。

Conversely, ScoreCAM demonstrates a more balanced approach by highlighting both osteogenic and osteoclastogenic regions, resulting in a more precise delineation of lesion boundaries. This distinction underscores the importance of selecting the appropriate methodology for image analysis in order to achieve optimal results in the identification and characterization of bone abnormalities.

相反，ScoreCAM通过突出显示成骨区域和破骨区域展示了一种更加平衡的方法，从而更精确地描绘了病灶边界。这一区别强调了选择适当的图像分析方法的重要性，以便在骨骼异常的识别和特征描述中取得最佳结果。

Further investigation into the comparative effectiveness of these techniques may yield valuable insights for enhancing diagnostic accuracy and treatment planning in the field of medical imaging. Additionally, a group of experienced radiologists is enlisted to provide insightful clinical explanations for instances of misjudgment in representative cases, thereby facilitating a deeper comprehension of the models’ functionality and ultimately improving its utility in the medical field..

进一步研究这些技术的相对有效性，可能会为提高医学影像领域中的诊断准确性和治疗规划提供宝贵的见解。此外，还召集了一组经验丰富的放射科医生，为典型案例中的误判情况提供有见地的临床解释，从而促进对模型功能的深入理解，并最终提升其在医学领域的实用性。

Manual annotations of ROI which served as ground truth for various deep learning models have long been regarded as a relatively challenging and intricate task, especially in CT- or MRI-based deep learning models

手动标注ROI作为各种深度学习模型的ground truth长期以来被认为是一个相对具有挑战性和复杂性的任务，尤其是在基于CT或MRI的深度学习模型中。

37

37

,

，

42

42

. Despite the continuous emergence of novel segmentation algorithms in recent years like Mask R-CNN, 3D CNN

. 尽管近年来不断涌现出像 Mask R-CNN、3D CNN 这样的新型分割算法

43

四十三

,

，

44

44

and so on, the segmentation performance of models built upon these algorithms often falls short of expectations. Issues such as misidentifying lesion locations or producing inaccurate segmentations frequently result in IoU and Dice scores that do not meet desired standards. Such discrepancies can introduce bias into subsequent classification model assessments and necessitate intricate manual verification and corrections in later stages.

等等，基于这些算法构建的模型在分割性能上往往不尽如人意。诸如错误识别病灶位置或生成不准确的分割等问题，经常导致IoU和Dice得分无法达到预期标准。这种差异可能会在后续分类模型评估中引入偏差，并需要在后期进行复杂的人工验证和修正。

Therefore, in terms of research design, compared with multitask deep learning framework, our research prioritizes the accuracy and interpretability of the deep learning model. All of the segmentation and labeling of lesion areas in the radiographs are meticulously carried out by professional radiologists..

因此，在研究设计上，与多任务深度学习框架相比，我们的研究更优先考虑深度学习模型的准确性和可解释性。所有关于放射影像中病变区域的分割和标记都由专业放射科医生精心完成。

The utilization of deep learning techniques has significantly improved the clinical diagnosis of medical images in computer-assisted imaging settings. Despite these advancements, distinguishing between PBTs and bone infections remains a challenging task. Previous research has successfully developed and validated deep learning models for classifying different types of PBTs using radiographic and demographic data.

深度学习技术的运用大大提高了计算机辅助成像环境中医学图像的临床诊断水平。尽管有了这些进步，区分PBTs和骨感染仍然是一项具有挑战性的任务。以往的研究已经成功开发并验证了使用X光片和人口统计数据对不同类型的PBTs进行分类的深度学习模型。

28

28

,

，

45

45

. However, these studies primarily concentrate on categorizing benign, intermediate, and malignant PBTs, rather than differentiating bone tumors from other musculoskeletal diseases that may be easily confused with PBTs. It is worth noting that while MRI-based deep learning models have been created to enhance the diagnosis of patients with PBTs and bone infections.

然而，这些研究主要集中在将良性、中间性和恶性PBTs分类，而不是将骨肿瘤与其他可能与PBTs混淆的肌肉骨骼疾病区分开来。值得注意的是，虽然已经创建了基于MRI的深度学习模型，以改善PBTs和骨感染患者的诊断。

42

42

, biases were present in the patient data collection due to variations in diagnosis and treatment protocols across different medical centers. Furthermore, these studies have overlooked important biomarkers such as CRP, ESR, ALP, lactate dehydrogenase (LDH) and so on. Combining the completeness of clinical information can better restore the original appearance and characteristics of the disease.

，由于不同医疗中心之间诊断和治疗方案的差异，患者数据收集中存在偏差。此外，这些研究忽略了重要的生物标志物，如C反应蛋白（CRP）、红细胞沉降率（ESR）、碱性磷酸酶（ALP）、乳酸脱氢酶（LDH）等。结合临床信息的完整性可以更好地还原疾病原始的表现和特征。

Our ensemble model which encompasses sufficient clinical information outperformed the other four models merely based on the image data. These cases underscore the necessity for more systematic approaches to data gathering and organization, encompassing a broader spectrum of bone lesions and data points to enhance the accuracy of the models..

我们的集成模型包含了足够的临床信息，其表现优于其他四个仅基于图像数据的模型。这些案例强调了需要更系统的方法来收集和组织数据，涵盖更广泛的骨病变和数据点，以提高模型的准确性。

This study has limitations. Firstly, bone infections are more common than PBTs and benign subtypes in PBTs are far more common than malignant ones. However, because the hospitals selected were regionally superior medical centers, patients with intractable diseases have high tendency. Secondly, our external validation set includes a children’s specialty hospital (Hospital 3), while it does help increase the diversity of our study population to some extent, making our research more representative, it may introduce some bias in terms of population distribution.

本研究存在一些局限性。首先，骨感染比PBTs更常见，且PBTs中的良性亚型远比恶性亚型常见。然而，由于所选医院为区域性优质医疗中心，疑难病患者比例较高。其次，我们的外部验证集包含了一家儿童专科医院（医院3），虽然这在一定程度上有助于增加研究人群的多样性，使我们的研究更具代表性，但在人群分布上可能引入了一些偏差。

Thirdly, the segmentation and labeling of lesion areas in the radiographs were entirely carried out by radiologists manually, making the research multifarious, although it may bring better work. In addition, in the collection process of clinical information, we found that for some examination like ALP and LDH, not all patients need this examination.

第三，放射影像中对病灶区域的分割和标记完全由放射科医生手动完成，这使得研究工作变得复杂，尽管这可能会带来更好的结果。此外，在临床信息的收集过程中，我们发现对于像ALP和LDH这样的检查，并不是所有患者都需要进行这项检查。

In addition, doctors from different hospitals and departments may also exist examination preference, which lead to large amount of missing information. In the future, more cases with radiograph images from representative hospitals and more standardized collection of clinical information need to be researched to improve the generalizability and completeness of the model..

此外，来自不同医院和科室的医生可能存在检查偏好，这导致了大量信息缺失。未来，需要研究更多来自代表性医院的含X光片的病例，并更规范地收集临床信息，以提高模型的泛化性和完整性。

This groundbreaking study introduces a radiograph-based deep learning framework designed to enhance the classification of PBTs and bone infections, while also elucidating the clinical interpretation of these models. The ensemble deep learning framework, utilizing multicenter radiographs and clinical data, significantly improves the diagnostic accuracy for the binary classification.

这项开创性研究介绍了一种基于放射影像的深度学习框架，旨在提高对PBT和骨感染分类的准确性，同时阐明这些模型的临床解释。该集成深度学习框架利用多中心放射影像和临床数据，显著提升了二元分类的诊断准确性。

The results of the model have been meticulously visualized and professionally explained by expert radiologists. The ensemble model is more accurate and reliable in diagnosis compared with radiologists. These findings hold immense potential to guide orthopedic surgeons in making informed treatment decisions, thereby facilitating timely interventions for patients in need..

该模型的结果已经被专业的放射科医生精心可视化并专业解释。与放射科医生相比，集成模型在诊断上更加准确和可靠。这些发现具有巨大的潜力，可以指导骨科医生做出明智的治疗决策，从而帮助需要的患者及时进行干预。

Methods

方法

In this research, the methodology is mainly composed of data collection, preprocessing, annotation, model design, and development. The subsequent analysis was performed in compliance with all relevant ethical regulations, including the Declaration of Helsinki, as approved by the institutional review board of human studies of the Second Xiangya Hospital of Central South University (protocol number: no.2022-040) (Hospital 1).

在这项研究中，方法主要包括数据收集、预处理、标注、模型设计与开发。后续分析遵循了所有相关的伦理规定，包括《赫尔辛基宣言》，并经中南大学湘雅二医院人体研究机构审查委员会批准（协议编号：2022-040）（医院1）。

In addition, this retrospective study was approved by the local institutional review boards of Xiangya Hospital of Central South University (Hospital 2) and Hunan Children’s Hospital of Central South University (Hospital 3), and informed consent was waived because of the retrospective nature.

此外，这项回顾性研究获得了中南大学湘雅医院（医院2）和中南大学湖南儿童医院（医院3）当地机构审查委员会的批准，并且由于研究的回顾性性质，豁免了知情同意。

28

28

. The study was performed in accordance with national and international guidelines, and followed the recommended guidelines Checklist for Artificial Intelligence in Medical Imaging (CLAIM) guidelines (Supplementary Table

该研究是根据国家和国际指南进行的，并遵循了医学影像人工智能推荐指南清单（CLAIM）指南（补充表）。

8

8

)

)

46

46

.

。

Research participants and data

研究参与者和数据

This retrospective multicenter study collected patients via consecutive sampling between 2013 and 2022 from two cohorts: training cohort (from Hospital 1) and testing cohort (from Hospital 2 and Hospital 3) (Supplementary Fig.

这项回顾性多中心研究通过连续采样收集了2013年至2022年间来自两个队列的患者：训练队列（来自医院1）和测试队列（来自医院2和医院3）（补充图）。

1

1

). After screening, 1569 patients diagnosed of PBTs or bone infections with histopathology reports available as reference were finally included in the internal dataset. While 423 patients from another two medical centers were collected for validation (Fig.

). 经过筛选，最终纳入内部数据集的有1569名经组织病理学报告确诊为PBTs或骨感染的患者。而来自另外两家医疗中心的423名患者被收集用于验证（图。

2a

2a

and Supplementary Fig.

和补充图。

1

1

). These lesions were identified to have bone involvement through preoperative radiographs and were histologically diagnosed following biopsy or surgery. The criteria for evaluating the accuracy of both expert classifications and model classifications are grounded in pathological results, serving as the “ground truth”.

这些病灶通过术前X光片确认有骨骼受累，并在活检或手术后经组织学诊断。评估专家分类和模型分类准确性的标准基于病理结果，作为“真实情况”。

(i) For the inclusion criteria, lesions were confirmed and diagnosed as PBTs according to the 2020 World Health Organization (WHO) system for the classification for tumors of bone.

(i) 对于纳入标准，根据2020年世界卫生组织（WHO）骨肿瘤分类系统，病灶被确认并诊断为原发性骨肿瘤（PBTs）。

1

1

while bone infections were confirmed and proven by histology and (or) bacterial culture. The other vital inclusion criteria are evident as well as available clinical information and preoperative radiographs. (ii) The screening criteria were respectively described in Fig.

虽然骨感染已通过组织学和（或）细菌培养得到确认和证实。其他重要的纳入标准也同样明显，且有可用的临床信息和术前X光片。（ii）筛选标准分别在图中描述。

2b

2b

: (a) radiographs were from patients diagnosed between 2013 and 2022 (b) in selected three hospitals; (c) radiographs with robust quality for reliable assessments of the bone lesions and (d) all of these radiographs were preoperative. With reference to previous literature

(a) 放射影像来自2013年至2022年之间确诊的患者；(b) 选自三家医院；(c) 影像质量良好，能够可靠评估骨病变；(d) 所有这些放射影像均为术前影像。参考以往文献

42

42

,

，

45

45

,

，

47

47

, clinical characteristics of the included patients’ contained age, gender, lesion position (appendicular or axial), “whether the lesion painful ?”, “whether the lesion swelling ?”, “whether a recent history of trauma ?”, and we further collected examination data including C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), and alkaline phosphatase (ALP).

，纳入患者的临床特征包括年龄、性别、病灶位置（四肢或轴向）、“病灶是否疼痛？”、“病灶是否肿胀？”、“近期是否有外伤史？”，并进一步收集了检查数据，包括C反应蛋白（CRP）、红细胞沉降率（ESR）和碱性磷酸酶（ALP）。

All of the clinical data of the patients were reviewed and obtained from the patients’ electronic medical records after data desensitization and standardization..

所有患者的临床数据均经过去隐私化和标准化处理后，从患者的电子病历中回顾和获取。

Image preprocessing and annotation

图像预处理和标注

During the preprocessing stage, all of the radiographs were screened and selected based on the inclusion and exclusion criteria above. Notably, radiograph images like artifacts or foreign bodies which might significantly hinder the observation of lesions were regarded as poor-quality radiographs. One senior seniority radiologist (Y.H.) with systematic musculoskeletal fellowship training (12 years work experience) and one medium seniority clinical orthopedist (C.T.) (8 years work experience) independently reviewed these radiographs without the patients’ information, and the quality of them would decide by consensus.

在预处理阶段，所有X光片均根据上述纳入和排除标准进行筛选和选择。值得注意的是，像伪影或异物等可能显著妨碍病灶观察的X光图像被视为低质量X光片。一位资深放射科医生（Y.H.，具有系统的肌肉骨骼专科培训背景，12年工作经验）和一位中等资深临床骨科医生（C.T.，8年工作经验）在不查看患者信息的情况下独立审查了这些X光片，并通过共识决定其质量。

Radiographs were kept and downloaded as Digital Imaging and Communications in Medicine (DICOM) files from the picture archiving and communication system (PACS) at their original sizes and resolutions. All of these radiograph images have undergone desensitization processing of disengaging patient-protected health information from DICOM data to meet the relevant legal criteria and requirements of US (HIPAA) as well as European (GDPR).

射线照片以原始大小和分辨率从图像存档和通信系统 (PACS) 中保存并下载为医学数字成像和通信 (DICOM) 文件。所有这些射线照片图像都经过了脱敏处理，从 DICOM 数据中剥离受保护的患者健康信息，以符合美国 (HIPAA) 和欧洲 (GDPR) 的相关法律标准和要求。

42

42

. Delineating the region of interest (ROI) was performed by two proficient radiologists (Y.Q. with 3-5 years of experience and J.G. with 3-5 years of experience in screening musculoskeletal radiographs images). ROIs were meticulously outlined via Click 2 Crop (version 5.2.2) (

. 利用两名经验丰富（Y.Q. 和 J.G. 均有3-5年筛查肌肉骨骼X光片经验）的放射科医生划定感兴趣区域（ROI）。ROI通过Click 2 Crop（版本5.2.2）精心勾画 (

https://click-2-crop.en.softonic.com/

https://click-2-crop.en.softonic.com/

) to closely segment pertinent entities present in each PBT or bone infection. Instances where disagreements arose between the two radiologists regarding contentious boundaries of these entities were subjected to further scrutiny. In such cases, a distinguished senior radiologist (Y.H.), boasting an impressive 12 years of experience in screening musculoskeletal radiographs, undertook the task of confirming the final delineations of ROIs.

）以密切分割存在于每个PBT或骨感染中的相关实体。两位放射科医生在这些实体的争议边界上出现分歧的情况下，会进行进一步的审查。在这种情况下，由一位拥有12年筛查肌肉骨骼X光片经验的资深高级放射科医生（Y.H.）负责确认感兴趣区域（ROI）的最终划分。

The smallest rectangular box that can completely cover the ROI was manually annotated as the boundary box by senior seniority radiologist (Y.H.) to ensure accuracy. Afterward, the annotated ROIs were used as ground truth for the model development process..

由资深放射科医生 (Y.H.) 手动将能够完全覆盖感兴趣区域 (ROI) 的最小矩形框标注为边界框，以确保准确性。随后，标注的 ROI 被用作模型开发过程的参考标准。

Design of the imaging models

成像模型的设计

For the classification of the radiographs, imaging models were built upon four distinct neural networks: EfficientNet B3 (E3), EfficientNet B4 (E4), Vision Transformer (ViT), and Swin Transformers (SWIN)

为了对射线照片进行分类，基于四个不同的神经网络构建了成像模型：EfficientNet B3（E3）、EfficientNet B4（E4）、Vision Transformer（ViT）和Swin Transformers（SWIN）。

48

48

,

，

49

49

,

，

50

50

. These models were selected based on their state-of-the-art performance in image classification tasks and their ability to capture diverse features from medical images. Specifically, EfficientNet represents a lineage of Convolutional Neural Networks (CNNs) that utilize compound scaling to harmonize the depth, width, and resolution of the network, achieving optimal performance with fewer parameters compared to traditional CNNs.

这些模型因其在图像分类任务中的先进性能以及从医学图像中捕捉多样特征的能力而被选中。具体来说，EfficientNet 代表了一系列卷积神经网络 (CNN)，利用复合缩放方法协调网络的深度、宽度和分辨率，与传统 CNN 相比，以更少的参数实现了最佳性能。

50

50

. Thanks to this innovative methodology, EfficientNet consistently attains state-of-the-art accuracy, yet with markedly fewer parameters. This makes it a prime choice for an array of computer vision applications

由于采用了这种创新的方法，EfficientNet始终能够达到最先进的准确性，但参数明显更少。这使其成为众多计算机视觉应用的首选。

50

50

,

，

51

51

. The Vision Transformer (ViT) introduces a novel architecture that processes images as sequences of patches using Transformer blocks, originally designed for natural language processing tasks. This architecture has demonstrated significant potential in handling visual data. The Swin Transformer further refines this approach by incorporating a hierarchical structure and local self-attention mechanisms, enabling it to manage diverse resolutions and scales effectively.

Vision Transformer（ViT）引入了一种新颖的架构，利用最初为自然语言处理任务设计的Transformer模块，将图像作为序列的片段进行处理。该架构在处理视觉数据方面展现了巨大的潜力。Swin Transformer通过引入分层结构和局部自注意力机制进一步完善了这一方法，使其能够有效地处理不同的分辨率和尺度。

Collectively, these models represent some of the most advanced frameworks in computer vision..

这些模型共同代表了计算机视觉中最先进的框架之一。

Addressing the constraints of our limited label data, we adopted a transfer learning strategy. All four imaging models were initialized with weights pre-trained on the extensive ImageNet dataset, followed by fine-tuning on our proprietary bone dataset

为了解决我们有限标签数据的限制，我们采用了迁移学习策略。所有四个成像模型均使用在大规模ImageNet数据集上预训练的权重进行初始化，随后在我们的专有骨骼数据集上进行微调。

52

52

. The original classification heads of these models, designed for 1000-class classification, were replaced with a single output node equipped with a sigmoid activation function to facilitate binary predictions (PBTs vs. bone infection).

这些模型的原始分类头设计用于1000类分类，被替换为带有sigmoid激活函数的单个输出节点，以实现二元预测（PBT与骨感染）。

Model training and evaluation

模型训练与评估

The internal dataset from Hospital 1 was partitioned into training, validation, and test set at a ratio of 7:1:2, respectively. The dataset from Hospital 2 and Hospital 3 was set aside as an external test set to evaluate the generalizability of our models across different data sources. Each of the four imaging models was trained independently using a batch size of 128 over 100 epochs.

医院1的内部数据集按照7:1:2的比例分别划分为训练集、验证集和测试集。医院2和医院3的数据集被留作外部测试集，以评估我们的模型在不同数据来源上的泛化能力。四个影像模型中的每一个都使用批量大小为128的数据独立训练了100个周期。

We employed Binary Cross-Entropy loss as our loss function. Optimization of the model was achieved through Stochastic Gradient Descent with an initial learning rate of 0.1. This rate was decayed by a factor of 10 every 30 epochs. For testing, we utilized the weights from the epoch exhibiting the best performance on the validation dataset..

我们使用二元交叉熵损失作为损失函数。模型的优化通过随机梯度下降实现，初始学习率为0.1。该学习率每30个周期按10倍衰减。在测试中，我们使用了在验证集上表现最佳的周期对应的权重。

Our algorithms were developed in Python 3.7 and executed on a machine equipped with an NVIDIA RTX 3090 GPU. The deep learning framework used in this study is PyTorch. In terms of data preprocessing, all images underwent resizing and normalization. Specifically, images were resized to a resolution of 224 × 224 pixels and normalized using the mean and standard deviation of the training dataset.

我们的算法使用 Python 3.7 开发，并在配备 NVIDIA RTX 3090 GPU 的机器上运行。本研究使用的深度学习框架是 PyTorch。在数据预处理方面，所有图像都进行了调整大小和归一化处理。具体而言，图像被调整为 224 × 224 像素的分辨率，并使用训练数据集的均值和标准差进行归一化。

To further enhance performance, we incorporated standard data augmentation techniques during training, including random horizontal and vertical flips with a probability of 0.5 for each..

为了进一步提升性能，我们在训练过程中加入了标准的数据增强技术，包括以0.5的概率进行随机水平和垂直翻转。

Model ensemble

模型集成

To further optimize performance, we integrated the predictions from the four imaging models (E3, E4, ViT, and SWIN) with traditional machine-learning models based on patients’ clinical characteristics. The hyperparameters utilized in the four imaging models and the ensemble model are depicted in Supplementary Table .

为了进一步优化性能，我们将四个影像模型（E3、E4、ViT 和 SWIN）的预测结果与基于患者临床特征的传统机器学习模型相结合。这四个影像模型和集成模型中使用的超参数展示在补充表中。

9

9

. Specifically, we designed and evaluated several machine-learning models, including Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosted Decision Trees (GBDT), Light Gradient Boosting Machine (LightGBM), Decision Tree (DT), Logistics Regression (LR), Extreme Gradient Boosting (XGBoost) and K-Nearest Neighbor (KNN).

具体来说，我们设计并评估了多种机器学习模型，包括随机森林（RF）、自适应提升（AdaBoost）、梯度提升决策树（GBDT）、轻量级梯度提升机（LightGBM）、决策树（DT）、逻辑回归（LR）、极端梯度提升（XGBoost）和K近邻（KNN）。

Given the missing clinical data and the significant differences in clinical features between PBTs and bone infections, the clinical characteristics included in the ensemble model were age, gender, and lesion location..

鉴于缺失的临床数据以及PBTs和骨感染之间临床特征的显著差异，集成模型中包含的临床特征为年龄、性别和病灶位置。

The construction of the ensemble model involved a two-step 5-fold cross-validation approach to avoid self-validation. In the first step, the four trained imaging models were used to score each patient (Supplementary Fig.

集成模型的构建涉及一个两步的5折交叉验证方法，以避免自我验证。在第一步中，使用四个训练好的成像模型对每位患者进行评分（补充图）。

5

5

). In the second step, these scores were integrated with clinical features using traditional machine-learning methods, with fivefold cross-validation utilized for hyperparameter tuning (Supplementary Fig.

）。在第二步中，这些评分通过传统机器学习方法与临床特征相结合，并使用五折交叉验证进行超参数调整（补充图。

6

6

). Through systematic comparison, we determined that the ensemble model utilizing Random Forest achieved the highest AUC (Supplementary Fig.

通过系统的比较，我们确定使用随机森林的集成模型达到了最高的AUC（补充图）。

7

7

). The final ensemble framework integrates both clinical characteristics and imaging information, providing a comprehensive diagnostic tool for PBTs and bone infection classification.

）。最终的集成框架结合了临床特征和影像信息，为PBTs和骨感染分类提供了一个全面的诊断工具。

Visualization and examples

可视化与示例

To interpret the models’ predictions, we use GradCAM and ScoreCAM to visualize the regions that our model relies on for decision-making. GradCAM calculates the gradient of the target class score with respect to feature maps. It then applies global-average-pooling to these gradients to determine the importance weights for each feature map.

为了诠释模型的预测，我们使用GradCAM和ScoreCAM来可视化我们的模型用于决策的区域。GradCAM计算目标类别得分相对于特征图的梯度。然后，它对这些梯度应用全局平均池化，以确定每个特征图的重要性权重。

This weighted combination, when subjected to a ReLU activation, produces a coarse localization map highlighting the most relevant image regions. As GradCAM is model-agnostic, it can be applied to four different models in our approach. In contrast, ScoreCAM, an extension of GradCAM, does not use gradients.

这种加权组合在经过 ReLU 激活后，会产生一个粗略的定位图，突出显示最相关的图像区域。由于 GradCAM 是模型无关的，因此可以应用于我们方法中的四种不同模型。相比之下，GradCAM 的扩展版本 ScoreCAM 不使用梯度。

Instead, it activates each feature map in the target layer individually and forwards these to obtain the class score. The final saliency map is derived by linearly combining the activation maps with their respective scores. This results in sharper and more precise visual explanations than GradCAM provides.

相反，它单独激活目标层中的每个特征图，并将其前向传播以获得类别得分。最终的显著性图通过将其激活图与其各自得分线性组合得出。这比 GradCAM 提供了更清晰、更精确的视觉解释。

Together, these two methods offer insights into the regions of an X-ray that our model considers essential for predictions..

这两种方法共同提供了对我们模型认为对预测至关重要的X射线区域的洞察。

Radiologist evaluation

放射科医生评估

To assess and contrast the precision of clinical doctors and the classification judgments made by various deep learning models, we have enlisted the participation of three distinct groups of radiologists varying in seniority. Within this study, three expert groups (EG) with different seniority were designed.

为了评估和对比临床医生的准确性以及各种深度学习模型的分类判断，我们邀请了三组不同资历的放射科医生参与。在本研究中，设计了三组不同资历的专家组（EG）。

Individuals classified as junior radiologists possessed 2–4 years of experience (Q.L. and J.G.) and were responsible for analyzing 1500 musculoskeletal radiograph reports annually (EG1). While senior radiologists (Prof. P. and Prof. L.) had accumulated over 10 years of experience in the field (EG3).

初级放射科医生（Q.L. 和 J.G.）拥有2至4年的工作经验，每年负责分析1500份肌肉骨骼X光报告（EG1）。而资深放射科医生（P.教授和L.教授）则在该领域积累了超过10年的经验（EG3）。

42

42

,

，

47

47

. In addition, we engaged another group of refresher radiologists (M.W. and Y.Z.) with 8–10 years of experience referred as medium seniority group (EG2). Each radiologist independently evaluated radiographs and associated clinical data using a conventional PACS system, with the diagnoses being made without prior knowledge of the pathological and/or bacterial culture results.

此外，我们还邀请了另一组具有8-10年经验的放射科医生（M.W.和Y.Z.），称为中资深组（EG2）。每位放射科医生独立使用传统的PACS系统评估了放射影像及其相关的临床数据，诊断过程中并不事先了解病理和/或细菌培养结果。

The inter-reader reliability among radiologists were evaluated through Fleiss .

放射科医生之间的阅片者间可靠性通过Fleiss方法进行了评估。

κ

κ

and Cohen

和科恩

κ

κ

53

53

.

。

Statistics analysis

统计分析

All statistical analyses were conducted using the opensource R software (version 4.2.3; R Foundation). Evaluation of the classification performance involved the use of the receiver operating characteristic (ROC) curve, along with metrics such as the area under the curve (AUC), accuracy, sensitivity, specificity, and confusion matrices.

所有统计分析均使用开源的R软件（版本4.2.3；R基金会）进行。分类性能的评估涉及接收者操作特征（ROC）曲线，以及曲线下面积（AUC）、准确性、敏感性、特异性和混淆矩阵等指标。

The mean AUC was specifically employed to assess the average performance of these four distinct imaging models. Statistical differences in clinicopathologic features among groups were analyzed using the Kruskal–Wallis rank-sum test for continuous variables and the chi-square test for categorical variables.

平均AUC被专门用于评估这四种不同成像模型的平均性能。组间临床病理特征的统计学差异使用Kruskal-Wallis秩和检验分析连续变量，使用卡方检验分析分类变量。

Statistical differences between the AUC curves of different models were assessed using the DeLong test.

使用DeLong检验评估不同模型的AUC曲线之间的统计学差异。

54

54

, while the statistical differences between the models and radiologist experts were evaluated using the Cochran’s Q test

，而模型与放射学专家之间的统计差异则通过Cochran's Q检验进行评估

55

55

,

，

56

56

, which is appropriate for multiple sets of paired data. Calculation of 95% confidence intervals (CI) was performed using the Wilson method.

适用于多组配对数据。使用Wilson方法计算95%置信区间（CI）。

P

P

values below 0.05 were considered as statistically significant.

低于0.05的值被认为具有统计学意义。

Data availability

数据可用性

The raw data collected and processed in this study are supervised under the corresponding institutions. All of the imaging data in this study has been desensitized and publicly released with restricted access on Zenodo (

本研究收集和处理的原始数据受相应机构监管。本研究中的所有影像数据均已脱敏，并在Zenodo上以受限访问的方式公开发布（

https://zenodo.org/

https://zenodo.org/

) at

) 在

https://doi.org/10.5281/zenodo.13858807

https://doi.org/10.5281/zenodo.13858807

. This DOI represents all versions, and will always resolve to the latest one. The data are available by emailing the corresponding author with all requests for academic use. The requirements will be evaluated concerning institutional policies, and data can only be shared for non-commercial academic usage with a formal material transfer agreement.

.该DOI代表所有版本，并将始终解析为最新版本。数据可通过电子邮件联系通讯作者获取，所有学术用途的请求均需提出。要求将根据机构政策进行评估，数据只能用于非商业学术用途，并需签订正式的材料转让协议。

All requests will be promptly reviewed within a timeframe of 30 working days..

所有请求将在30个工作日内及时审查。

Code availability

代码可用性

The pipeline development and experiments are conducted in Python with PyTorch as a primary tool. All of the codes for reproducing this study (Deep learning pipeline for Tumors and Infections based on Radiographs Predicition) can be found at

管道开发和实验是使用 Python 进行的，PyTorch 作为主要工具。本研究（基于射线照片预测的肿瘤和感染深度学习管道）的所有代码均可在此处找到：

https://github.com/CSUXY-2YY/DeepTIRP

https://github.com/CSUXY-2YY/DeepTIRP

.

。

References

参考文献

Choi, J. H. & Ro, J. Y. The 2020 WHO classification of tumors of bone: an updated review.

崔，J. H. & 卢，J. Y. 2020年WHO骨肿瘤分类：更新回顾。

Adv. Anat. Pathol.

解剖病理学进展

28

28

, 119–138 (2021).

，119-138（2021）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Ferguson, J. L. & Turner, S. P. Bone cancer: diagnosis and treatment principles.

弗格森，J. L. 和 特纳，S. P. 骨癌：诊断与治疗原则。

Am. Fam. Physician

美国家庭医生

98

98

, 205–213 (2018).

，205-213页（2018年）。

PubMed

PubMed

Google Scholar

谷歌学术索

Molina, E. R., Chim, L. K., Barrios, S., Ludwig, J. A. & Mikos, A. G. Modeling the tumor microenvironment and pathogenic signaling in bone sarcoma.

莫利纳，E. R.，钱姆，L. K.，巴里奥斯，S.，路德维希，J. A.，米科斯，A. G. 骨肉瘤中的肿瘤微环境和致病信号建模。

Tissue Eng. Part B Rev.

组织工程学。 B部分 评论。

26

26

, 249–271 (2020).

，249-271页（2020年）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Siegel, R. L., Miller, K. D., Wagle, N. S. & Jemal, A. Cancer statistics, 2023.

西格尔，R. L.，米勒，K. D.，瓦格尔，N. S.，杰马尔，A. 癌症统计，2023。

CA Cancer J. Clin.

CA 癌症临床期刊

73

73

, 17–48 (2023).

，17-48页（2023年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Bölling, T., Hardes, J. & Dirksen, U. Management of bone tumours in paediatric oncology.

Bölling, T., Hardes, J. & Dirksen, U. 儿童肿瘤学中骨肿瘤的管理。

Clin. Oncol. (R. Coll. Radiol.)

临床肿瘤学（皇家放射学院）

25

25

, 19–26 (2013).

，19-26页（2013年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Zhang, W. et al. PRKDC induces chemoresistance in osteosarcoma by recruiting GDE2 to stabilize GNAS and activate AKT.

张伟等。PRKDC通过招募GDE2来稳定GNAS并激活AKT，从而诱导骨肉瘤的化疗耐药性。

Cancer Res

癌症研究

84

84

, 2873–2887 (2024).

，2873–2887（2024）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Yu, S. & Yao, X. Advances on immunotherapy for osteosarcoma.

余珊、姚欣。骨肉瘤免疫治疗的进展。

Mol. Cancer

分子癌症

23

23

, 192 (2024).

，192（2024）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术搜索

Casali, P. G. et al. Bone sarcomas: ESMO-PaedCan-EURACAN Clinical Practice Guidelines for diagnosis, treatment and follow-up.

卡萨利，P. G. 等。骨肉瘤：ESMO-PaedCan-EURACAN 临床实践指南，用于诊断、治疗和随访。

Ann. Oncol.

安·昂科尔。

29

29

, iv79–iv95 (2018).

，iv79–iv95（2018）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Liu, J. et al. Anticancer and bone-enhanced nano-hydroxyapatite/gelatin/polylactic acid fibrous membrane with dual drug delivery and sequential release for osteosarcoma.

刘, J. 等。具有双重药物递送和序贯释放的抗癌和骨增强纳米羟基磷灰石/明胶/聚乳酸纤维膜用于骨肉瘤治疗。

Int. J. Biol. Macromol.

国际生物大分子杂志

240

240

, 124406 (2023).

，124406（2023）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术索

Meltzer, P. S. & Helman, L. J. New horizons in the treatment of osteosarcoma.

Meltzer, P. S. 和 Helman, L. J. 骨肉瘤治疗的新视野。

New Engl. J. Med.

新英格兰医学杂志

385

385

, 2066–2076 (2021).

，2066-2076（2021）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Li, J. et al. Primary bone tumor detection and classification in full-field bone radiographs via YOLO deep learning model.

李, J. 等。通过YOLO深度学习模型在全视野骨X光片中检测和分类原发性骨肿瘤。

Eur. Radiol.

欧洲放射学杂志

33

33

, 4237–4248 (2023).

，4237–4248页（2023年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Caracciolo, J. T., Temple, H. T., Letson, G. D. & Kransdorf, M. J. A modified lodwick-madewell grading system for the evaluation of lytic bone lesions.

卡拉乔洛，J.T.，坦普尔，H.T.，莱特森，G.D.，克兰斯多夫，M.J. 一种改良的Lodwick-Madewell溶骨性骨病变评估分级系统。

AJR Am. J. Roentgenol.

AJR 美国放射学杂志

207

207

, 150–156 (2016).

，150-156页（2016年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术搜索

Kovacs, S. K., Manassaporn, A., Nielsen, G. P. & Hung, Y. P. Molecular and immunohistochemical testing of bone tumours: review and update.

科瓦克斯，S. K.，马纳萨蓬，A.，尼尔森，G. P.，洪，Y. P. 骨肿瘤的分子和免疫组织化学检测：综述与更新。

Histopathology

组织病理学

82

82

, 794–811 (2023).

，794-811（2023）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Tao, Y. et al. Qualitative histopathological classification of primary bone tumors using deep learning: a pilot study.

陶, Y. 等。使用深度学习对原发性骨肿瘤进行定性组织病理学分类：一项初步研究。

Front. Oncol.

前沿. 肿瘤学

11

11

, 735739 (2021).

，735739（2021）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Rozeman, L. B., Cleton-Jansen, A. M. & Hogendoorn, P. C. Pathology of primary malignant bone and cartilage tumours.

罗泽曼, L. B., 克莱顿-扬森, A. M. & 霍亨杜恩, P. C. 原发性恶性骨与软骨肿瘤的病理学。

Int. Orthop.

国际骨科

30

30

, 437–444 (2006).

，第437-444页（2006年）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Kellish, A. S. et al. Reliability and accuracy in radiographic measurements of musculoskeletal tumors.

凯利什，A.S. 等。肌肉骨骼肿瘤放射测量的可靠性和准确性。

J. Orthop. Res.

骨科研究杂志

40

40

, 1654–1660 (2022).

，1654-1660（2022）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术索

Ulaner, G., Hwang, S., Landa, J., Lefkowitz, R. A. & Panicek, D. M. Musculoskeletal tumours and tumour-like conditions: common and avoidable pitfalls at imaging in patients with known or suspected cancer: Part B: malignant mimics of benign tumours.

乌兰纳，G.，黄，S.，兰达，J.，莱夫科维茨，R. A.，帕尼切克，D. M. 肌肉骨骼肿瘤及类似肿瘤的病变：在已知或疑似癌症患者的影像学中常见且可避免的陷阱：B部分：良性肿瘤的恶性模拟。

Int. Orthop.

国际骨科

37

37

, 877–882 (2013).

，877-882页（2013年）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Ulaner, G., Hwang, S., Lefkowitz, R. A., Landa, J. & Panicek, D. M. Musculoskeletal tumors and tumor-like conditions: common and avoidable pitfalls at imaging in patients with known or suspected cancer: Part A: benign conditions that may mimic malignancy.

乌兰纳，G.，黄，S.，莱夫科维茨，R. A.，兰达，J.，& 帕尼切克，D. M. 肌肉骨骼肿瘤和类肿瘤样病变：在已知或疑似癌症患者影像学中常见且可避免的陷阱：A部分：可能模拟恶性肿瘤的良性病变。

Int. Orthop.

国际骨科

37

37

, 871–876 (2013).

，871-876页（2013年）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Anderson, M. E., Wu, J. S. & Vargas, S. O. CORR (®) tumor board: do orthopaedic oncologists agree on the diagnosis and treatment of cartilage tumors of the appendicular skeleton?

安德森，M. E.，吴，J. S.，& 瓦尔加斯，S. O. CORR (®) 肿瘤委员会：骨科肿瘤医生是否对四肢骨骼软骨肿瘤的诊断和治疗达成一致？

Clin. Orthop. Relat. Res.

临床骨科及相关研究

475

475

, 2172–2175 (2017).

，2172–2175（2017）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Benz, M. R., Crompton, J. G. & Harder, D. PET/CT variants and pitfalls in bone and soft tissue sarcoma.

Benz, M. R., Crompton, J. G. & Harder, D. 骨与软组织肉瘤的PET/CT变异与陷阱。

Semin. Nucl. Med.

核医学研讨会

51

51

, 584–592 (2021).

，584-592页（2021年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Rozenberg, A. et al. Clinical impact of second-opinion musculoskeletal subspecialty interpretations during a multidisciplinary orthopedic oncology conference.

罗森伯格，A. 等。在多学科骨科肿瘤会议期间第二意见肌肉骨骼亚专科解读的临床影响。

J. Am. Coll. Radiol.

美国放射学会杂志

14

14

, 931–936 (2017).

，931-936页（2017年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Rozenberg, A. et al. Second opinions in orthopedic oncology imaging: can fellowship training reduce clinically significant discrepancies?

罗森伯格，A. 等。骨科肿瘤影像学的第二意见：专科培训能否减少临床显著差异？

Skelet. Radiol.

骨骼放射学。

48

48

, 143–147 (2019).

，143-147页（2019年）。

Article

文章

Google Scholar

谷歌学术

Mullowney, M. W. et al. Artificial intelligence for natural product drug discovery.

Mullowney, M. W. 等。用于天然产物药物发现的人工智能。

Nat. Rev. Drug Discov.

自然综述药物发现

22

22

, 895–916 (2023).

，895-916页（2023年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence.

Topol, E. J. 高性能医学：人类与人工智能的融合。

Nat. Med.

自然医学

25

25

, 44–56 (2019).

，44-56页（2019年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Zhu, W., Xie, L., Han, J. & Guo, X. The application of deep learning in cancer prognosis prediction.

朱伟，谢林，韩晶，郭霞。深度学习在癌症预后预测中的应用。

Cancers

癌症

12

12

, 603 (2020).

，603页（2020年）。

Chen, X. et al. Recent advances and clinical applications of deep learning in medical image analysis.

陈，X. 等。深度学习在医学图像分析中的最新进展及临床应用。

Med. Image Anal.

医学图像分析

79

79

, 102444 (2022).

，102444（2022）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术搜索

Painuli, D., Bhardwaj, S. & Köse, U. Recent advancement in cancer diagnosis using machine learning and deep learning techniques: a comprehensive review.

Painuli, D., Bhardwaj, S. & Köse, U. 使用机器学习和深度学习技术在癌症诊断中的最新进展：全面综述。

Comput Biol. Med.

计算生物学与医学。

146

146

, 105580 (2022).

，105580（2022）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

von Schacky, C. E. et al. Multitask deep learning for segmentation and classification of primary bone tumors on radiographs.

von Schacky, C. E. 等。多任务深度学习在X光片上对原发性骨肿瘤的分割与分类。

Radiology

放射学

301

301

, 398–406 (2021).

，398-406页（2021年）。

Article

文章

Google Scholar

谷歌学术

Murphey, M. D. et al. From the archives of AFIP. Imaging of giant cell tumor and giant cell reparative granuloma of bone: radiologic-pathologic correlation.

Murphey, M. D. 等。来自AFIP档案。骨巨细胞瘤和巨细胞修复性肉芽肿的影像学：放射学与病理学相关性。

Radiographics

放射图像学

21

21

, 1283–1309 (2001).

，1283-1309（2001）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术搜索

Miller, T. T. Bone tumors and tumorlike conditions: analysis with conventional radiography.

米勒，T.T. 骨肿瘤和类肿瘤样病变：常规X线摄影分析。

Radiology

放射学

246

246

, 662–674 (2008).

，662-674页（2008年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Delorme, S. & Baur-Melnyk, A. Imaging in multiple myeloma.

德尔莫尔，S. & 鲍尔-梅尔尼克，A. 多发性骨髓瘤的影像学。

Eur. J. Radiol.

欧洲放射学杂志

70

70

, 401–408 (2009).

，401-408页（2009年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

van de Meent, M. M., Pichardo, S. E. C., Rodrigues, M. F., Verbist, B. M. & van Merkesteyn, J. P. R. Radiographic characteristics of chronic diffuse sclerosing osteomyelitis/tendoperiostitis of the mandible: a comparison with chronic suppurative osteomyelitis and osteoradionecrosis.

van de Meent, M. M., Pichardo, S. E. C., Rodrigues, M. F., Verbist, B. M. & van Merkesteyn, J. P. R. 下颌骨慢性弥漫性硬化性骨髓炎/腱膜炎的放射学特征：与慢性化脓性骨髓炎和放射性骨坏死的比较。

J. Craniomaxillofac. Surg.

《颅面外科杂志》

46

46

, 1631–1636 (2018).

，1631–1636（2018）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术索

Gould, C. F., Ly, J. Q., Lattin, G. E. Jr., Beall, D. P. & Sutcliffe, J. B. 3rd Bone tumor mimics: avoiding misdiagnosis.

古尔德，C. F.，李，J. Q.，拉廷，G. E. Jr.，比尔，D. P.，萨特克利夫，J. B. 3rd 骨肿瘤模拟：避免误诊。

Curr. Probl. Diagn. Radiol.

当前诊断放射学问题

36

36

, 124–141 (2007).

，124-141页（2007年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Hill, B. G., Krogue, J. D., Jevsevar, D. S. & Schilling, P. L. Deep learning and imaging for the orthopaedic surgeon: how machines “read” radiographs.

希尔，B. G.，克罗格，J. D.，杰夫塞瓦，D. S.，施林，P. L. 深度学习和影像技术在骨科医生中的应用：机器如何“解读”X光片。

J. Bone Jt. Surg. Am.

美国骨科医师学会期刊

104

104

, 1675–1686 (2022).

，1675-1686（2022）。

Article

文章

Google Scholar

谷歌学术

Kijowski, R., Liu, F., Caliva, F. & Pedoia, V. Deep learning for lesion detection, progression, and prediction of musculoskeletal disease.

基霍夫斯基，刘，卡利瓦，佩多亚。深度学习在肌肉骨骼疾病病灶检测、进展和预测中的应用。

J. Magn. Reson Imaging

磁共振成像杂志

52

52

, 1607–1619 (2020).

，1607-1619（2020）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术搜索

Liu, P. et al. Deep learning to segment pelvic bones: large-scale CT datasets and baseline models.

刘, P. 等。深度学习分割骨盆骨骼：大规模CT数据集与基线模型。

Int. J. Comput. Assist Radiol. Surg.

计算机辅助放射学与外科国际期刊

16

16

, 749–756 (2021).

，749-756页（2021年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Noguchi, S. et al. Deep learning-based algorithm improved radiologists’ performance in bone metastases detection on CT.

野口，S. 等。基于深度学习的算法提高了放射科医生在CT上检测骨转移的能力。

Eur. Radiol.

欧洲放射学杂志

32

32

, 7976–7987 (2022).

，7976-7987（2022）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Arthur, A. et al. A CT-based radiomics classification model for the prediction of histological type and tumour grade in retroperitoneal sarcoma (RADSARC-R): a retrospective multicohort analysis.

Arthur, A. 等。基于CT的放射组学分类模型用于预测腹膜后肉瘤（RADSARC-R）的组织学类型和肿瘤分级：一项回顾性多队列分析。

Lancet Oncol.

柳叶刀肿瘤学

24

24

, 1277–1286 (2023).

，1277–1286（2023）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Hallinan, J. et al. Deep learning model for automated detection and classification of central canal, lateral recess, and neural foraminal stenosis at lumbar spine MRI.

Hallinan, J. 等。深度学习模型用于腰椎MRI中央管、侧隐窝和神经孔狭窄的自动检测与分类。

Radiology

放射学

300

300

, 130–138 (2021).

，130-138页（2021年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Wennmann, M. et al. Combining deep learning and radiomics for automated, objective, comprehensive bone marrow characterization from whole-body MRI: a multicentric feasibility study.

Wennmann, M. 等。结合深度学习和影像组学实现从全身MRI中自动、客观、全面的骨髓特征分析：一项多中心可行性研究。

Invest. Radiol.

放射学投资。

57

57

, 752–763 (2022).

，752-763页（2022年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Zheng, H. D. et al. Deep learning-based high-accuracy quantitation for lumbar intervertebral disc degeneration from MRI.

郑，H. D. 等。基于深度学习的MRI腰椎间盘退变高精度定量分析。

Nat. Commun.

自然通讯

13

13

, 841 (2022).

，841页（2022年）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Ye, Q. et al. Automatic detection, segmentation, and classification of primary bone tumors and bone infections using an ensemble multi-task deep learning framework on multi-parametric MRIs: a multi-center study.

叶, Q. 等。使用多参数MRI上的集成多任务深度学习框架自动检测、分割和分类原发性骨肿瘤和骨感染：一项多中心研究。

Eur. Radiol

欧洲放射学

.

。

34

34

, 4287–4299 (2023).

，4287-4299（2023）。

Bitarafan, A., Nikdan, M. & Baghshah, M. S. 3D image segmentation with sparse annotation by self-training and internal registration.

Bitarafan, A., Nikdan, M. & Baghshah, M. S. 通过自训练和内部配准实现稀疏标注的3D图像分割。

IEEE J. Biomed. Health Inf.

IEEE生物医学与健康信息学杂志

25

25

, 2665–2672 (2021).

，2665–2672（2021）。

Article

文章

Google Scholar

谷歌学术

Pereira, H. M., Marchiori, E. & Severo, A. Magnetic resonance imaging aspects of giant-cell tumours of bone.

佩雷拉，H. M.，马希奥里，E.，& 塞韦罗，A. 骨巨细胞瘤的磁共振成像表现。

J. Med. Imaging Radiat. Oncol.

医学影像与放射肿瘤学杂志

58

58

, 674–678 (2014).

，674-678页（2014年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

He, Y. et al. Deep learning-based classification of primary bone tumors on radiographs: a preliminary study.

何宇等。基于深度学习的X光片原发性骨肿瘤分类：初步研究。

EBioMedicine

EBioMedicine

62

62

, 103121 (2020).

，103121（2020）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Mongan, J., Moy, L. & Kahn, C. E. Jr. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers.

蒙根，J.，莫伊，L. & 卡恩，C. E. Jr. 医学影像人工智能检查清单（CLAIM）：作者和评审员指南。

Radiol. Artif. Intell.

放射学。人工智能。

2

2

, e200029 (2020).

，e200029（2020）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Eweje, F. R. et al. Deep learning for classification of bone lesions on routine MRI.

Eweje, F. R. 等。深度学习在常规MRI骨病变分类中的应用。

EBioMedicine

EBio医学

68

68

, 103402 (2021).

，103402（2021）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Heidari, M. et al. Enhancing efficiency in vision transformer networks: design techniques and insights. Preprint at

Heidari, M. 等。提升视觉Transformer网络的效率：设计技术与见解。预印本 tại

https://arxiv.org/abs/2403.19882

https://arxiv.org/abs/2403.19882

(2024).

（2024）。

Liu, Z. & et al. Swin transformer: hierarchical vision transformer using shifted windows. In

刘，Z. 等。Swin Transformer：使用移位窗口的分层视觉Transformer。在

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

IEEE/CVF国际计算机视觉会议（ICCV）论文集

10006-10017 (IEEE, 2021).

10006-10017（IEEE，2021）。

Tan, M. & Le, Q. V. EfficientNet: rethinking model scaling for convolutional neural networks. In

Tan, M. & Le, Q. V. EfficientNet：重新思考卷积神经网络的模型缩放。在

Proceedings of the 36th International Conference on Machine Learning

第36届国际机器学习会议论文集

, Vol. 97 (eds. Kamalika, C. & Ruslan, S.) 6105--6114 (PMLR, Proceedings of Machine Learning Research, 2019).

，第97卷（编辑：Kamalika, C. & Ruslan, S.）6105-6114页（PMLR，机器学习研究进展，2019）。

Mozaffari, J., Amirkhani, A. & Shokouhi, S. B. A survey on deep learning models for detection of COVID-19.

莫扎法里，J.，阿米尔霍萨尼，A.，肖库希，S. B. 深度学习模型在新冠病毒检测中的应用综述。

Neural Comput. Appl.

神经计算与应用

1–29 (2023).

1–29 (2023).

Deng, J. et al. ImageNet: a large-scale hierarchical image database. In

邓，J. 等。ImageNet：一个大规模的层次化图像数据库。在

2009 IEEE Conference on Computer Vision and Pattern Recognition

2009年IEEE计算机视觉与模式识别会议

248–255 (IEEE, 2009).

248–255 （IEEE，2009）。

Ragab, H. et al. DeePSC: a deep learning model for automated diagnosis of primary sclerosing cholangitis at two-dimensional MR cholangiopancreatography.

Ragab, H. 等。DeePSC：一种用于在二维磁共振胆胰管成像中自动诊断原发性硬化性胆管炎的深度学习模型。

Radiol. Artif. Intell.

放射学。人工智能。

5

5

, e220160 (2023).

，e220160（2023）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Kato, M. et al. A machine learning model for predicting the lymph node metastasis of early gastric cancer not meeting the endoscopic curability criteria.

加藤，M. 等。一种用于预测不符合内镜治愈标准的早期胃癌淋巴结转移的机器学习模型。

Gastric Cancer

胃癌

27

27

, 1069–1077 (2024).

，1069–1077（2024）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术搜索

Hou, P. et al. A paradigm shift in oncology imaging: a prospective cross-sectional study to assess low-dose deep learning image reconstruction versus standard-dose iterative reconstruction for comprehensive lesion detection in dual-energy computed tomography.

侯，P. 等。肿瘤学影像的范式转变：一项前瞻性横断面研究，评估低剂量深度学习图像重建与标准剂量迭代重建在双能量计算机断层扫描中对全面病灶检测的效果。

Quant. Imaging Med. Surg.

定量成像医学与外科

14

14

, 6449–6465 (2024).

，6449–6465（2024）。

Article

文章

PubMed

PubMed

PubMed Central

PubMed Central

Google Scholar

谷歌学术

Cochran, W. G. The comparison of percentages in matched samples.

科克伦，W. G. 配对样本中百分比的比较。

Biometrika

生物统计学

37

37

, 256–266 (1950).

，256-266页（1950年）。

Article

文章

PubMed

PubMed

Google Scholar

谷歌学术

Download references

下载参考文献

Acknowledgements

致谢

The authors would like to express our gratitude to BioRender (

作者要表达我们对BioRender的感谢（

https://app.biorender.com/

https://app.biorender.com/

) for assistance in creating the figures (Figs.

）感谢在创建图表（图

1

1

,

，

2

2

,

，

5

5

, and

，以及

6

6

). The authors are very grateful for the active participation of radiologists with diverse seniority: junior radiologist group (Q.L. and J.G.); medium seniority group (M.W. and Y.Z.); senior radiologist group (Prof. P. and Prof. L.). This work was supported by the National Natural Foundation of China (82272664, 82172500 and 32300528), The Science and Technology Innovation Program of Hunan Province (2023RC3085, 2023RC3080), Hunan Provincial Health High-Level Talent Scientific Research Project (R2023054), Hunan Provincial Natural Science Foundation of China (2022JJ30843), the Science and Technology Development Fund Guided by Central Government (2021Szvup169), Hunan Provincial Administration of Traditional Chinese Medicine Project (D2022117), Hunan Provincial Health High-Level Talent Scientific Research Project (R2023054), Key Project of Scientific Research of the Education Department of Hunan Province (24A0008), the Excellent Youth Foundation of Hunan Scientific Committee (2024JJ2084), the Scientific Research Fund of Hunan Provincial Education Department (23B0023) and the Scientific Research Program of Hunan Provincial Health Commission (B202304077077).

). 作者们对具有不同资历的放射科医生的积极参与表示由衷感谢：初级放射科医生组（Q.L. 和 J.G.）；中级资历组（M.W. 和 Y.Z.）；高级放射科医生组（P教授和L教授）。本研究得到了中国国家自然科学基金（82272664、82172500 和 32300528）、湖南省科技创新计划（2023RC3085、2023RC3080）、湖南省卫生健康高层次人才科研项目（R2023054）、中国湖南省自然科学基金（2022JJ30843）、中央政府引导的科技发展基金（2021Szvup169）、湖南省中医药管理局项目（D2022117）、湖南省教育厅重点科研项目（24A0008）、湖南省科学委员会优秀青年基金（2024JJ2084）、湖南省教育厅科研基金（23B0023）以及湖南省卫生健康委员会科研计划（B202304077077）的支持。

The study sponsors did not have any role in the study design, the collection, analysis and interpretation of data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication..

研究资助者在研究设计、数据的收集、分析和解释；手稿的准备、审查或批准；以及决定将手稿提交出版方面没有任何作用。

Author information

作者信息

Author notes

作者笔记

These authors contributed equally: Hua Wang, Yu He.

这些作者贡献相同：王华，何宇。

Authors and Affiliations

作者与所属机构

Department of Orthopaedics, The Second Xiangya Hospital of Central South University, Changsha, Hunan, China

中南大学湘雅二医院骨科，中国湖南长沙

Hua Wang, Lu Wan, Chenbei Li, Zhaoqi Li, Zhihong Li, Haodong Xu & Chao Tu

王华、万璐、李晨贝、李照琪、李志宏、徐浩东、涂超

Hunan Key Laboratory of Tumor Models and Individualized Medicine, The Second Xiangya Hospital of Central South University, Changsha, Hunan, China

湖南省肿瘤模型与个体化医学重点实验室，中南大学湘雅二医院，长沙，湖南，中国

Hua Wang, Lu Wan, Chenbei Li, Zhaoqi Li, Zhihong Li, Haodong Xu & Chao Tu

王华、万璐、李晨贝、李照琪、李志红、徐浩东、涂超

Department of Radiology, The Second Xiangya Hospital of Central South University, Changsha, Hunan, China

中南大学湘雅二医院放射科，中国湖南长沙

Yu He

何雨

Shenzhen Research Institute of Central South University, Guangdong, China

中国广东省中南大学深圳研究院

Zhihong Li & Chao Tu

李志宏 和 涂超

Center for Precision Health, McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, TX, USA

德克萨斯大学休斯顿健康科学中心，麦考密克生物医学信息学院，精准健康中心，休斯顿，德克萨斯州，美国

Haodong Xu

许浩东

Changsha Medical University, Changsha, Hunan, China

中国湖南省长沙市长沙医学院

Chao Tu

超图

Authors

作者

Hua Wang

王华

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术搜索

Yu He

何雨

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术索

Lu Wan

卢万

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术

Chenbei Li

李晨贝

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术

Zhaoqi Li

李昭琦

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术

Zhihong Li

李志宏

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术搜索

Haodong Xu

徐浩东

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索该作者在

PubMed

PubMed

Google Scholar

谷歌学术

Chao Tu

超图

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

PubMed

Google Scholar

谷歌学术索

Contributions

贡献

C.T. and H.D.X. conceived and designed the study, performed the data analysis. H.W. and Y.H. contributed to the data collection, results interpretation, and manuscript preparation. L.W., C.B.L. and Z.Q.L. were participated in data collection. Z.H.L. was responsible for the supervision of the project.

C.T. 和 H.D.X. 构思并设计了研究，进行了数据分析。H.W. 和 Y.H. 参与了数据收集、结果解释和手稿准备。L.W.、C.B.L. 和 Z.Q.L. 参与了数据收集。Z.H.L. 负责项目的监督工作。

All authors read and approved the final manuscript..

所有作者阅读并批准了最终手稿。

Corresponding authors

通讯作者

Correspondence to

致信给

Haodong Xu

许浩东

or

或

Chao Tu

超图

.

。

Ethics declarations

伦理声明

Competing interests

竞争利益

The authors declare no competing interests.

作者声明不存在竞争性利益。

Additional information

附加信息

Publisher’s note

出版商说明

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Springer Nature 对已发布地图中的管辖权声明和机构隶属关系保持中立。

Supplementary information

补充信息

Supplementary Information

补充信息

All loaded pretrained models

所有已加载的预训练模型

Rights and permissions

权利与许可

Open Access

开放获取

This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material.

本文根据知识共享署名-非商业性使用-禁止演绎 4.0 国际许可协议获得许可，该协议允许您在任何媒介或格式中进行非商业性的使用、分享、分发和复制，只要您对原作者和来源给予适当的署名，提供指向知识共享许可协议的链接，并说明是否对授权材料进行了修改即可。

You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

根据本许可证，您无权分享从本文或其部分内容改编的材料。本文中的图像或其他第三方材料包含在文章的Creative Commons许可证中，除非材料的署名行另有说明。如果材料未包含在文章的Creative Commons许可证中，并且您的预期用途不被法律规定允许或超出了允许的使用范围，您需要直接从版权持有人处获得许可。

To view a copy of this licence, visit .

要查看此许可证的副本，请访问 。

http://creativecommons.org/licenses/by-nc-nd/4.0/

http://creativecommons.org/licenses/by-nc-nd/4.0/

.

。

Reprints and permissions

重印和权限

About this article

关于本文

Cite this article

引用这篇文章

Wang, H., He, Y., Wan, L.

王，何，万

et al.

等。

Deep learning models in classifying primary bone tumors and bone infections based on radiographs.

基于X光片的深度学习模型在分类原发性骨肿瘤和骨感染中的应用。

npj Precis. Onc.

npj 精准肿瘤学

9

9

, 72 (2025). https://doi.org/10.1038/s41698-025-00855-3

，72（2025）。https://doi.org/10.1038/s41698-025-00855-3

Download citation

下载引用

Received

已收到

:

：

01 August 2024

2024年8月1日

Accepted

已接受

:

：

25 February 2025

2025年2月25日

Published

已发布

:

：

13 March 2025

2025年3月13日

DOI

数字对象标识符

:

：

https://doi.org/10.1038/s41698-025-00855-3

https://doi.org/10.1038/s41698-025-00855-3

Share this article

分享这篇文章

Anyone you share the following link with will be able to read this content:

任何你分享以下链接的人都将能够阅读此内容：

Get shareable link

获取可共享链接

Sorry, a shareable link is not currently available for this article.

抱歉，此文章目前无法提供可共享的链接。

Copy to clipboard

复制到剪贴板

Provided by the Springer Nature SharedIt content-sharing initiative

由 Springer Nature SharedIt 内容共享计划提供

Subjects

主题

Bone cancer

骨癌

Tumour virus infections

肿瘤病毒感染