AbstractAccurate treatment response assessment using serial CT scans is essential in oncological clinical trials. However, oncologists’ assessment following the Response Evaluation Criteria in Solid Tumors (RECIST) guideline is subjective, time-consuming, and sometimes fallible. Advanced liver cancer often presents multifocal hepatic lesions on CT imaging, making accurate characterization more challenging than with other malignancies.

摘要在肿瘤临床试验中，使用连续CT扫描进行准确的治疗反应评估至关重要。然而，肿瘤学家根据实体瘤反应评估标准（RECIST）指南进行的评估是主观的，耗时的，有时甚至容易出错。晚期肝癌在CT成像中常表现为多灶性肝脏病变，因此准确表征比其他恶性肿瘤更具挑战性。

In this work, we developed a tumor volume guided comprehensive objective response evaluation based on deep learning (RECORD) for liver cancer. RECORD performs liver tumor segmentation, followed by sum of the volume (SOV)-based treatment response classification and new lesion assessment. Then, it can provide treatment evaluations of response, stability, and progression, and calculates progression-free survival (PFS) and response time.

在这项工作中，我们开发了一种基于肝癌深度学习（RECORD）的肿瘤体积引导的综合客观反应评估。记录执行肝脏肿瘤分割，然后进行基于体积总和（SOV）的治疗反应分类和新病变评估。然后，它可以提供反应，稳定性和进展的治疗评估，并计算无进展生存期（PFS）和反应时间。

The RECORD pipeline was developed with both CNN and ViT backbones. Its performance was evaluated in three longitudinal cohorts involving 60 multi-national centers, 206 patients, 891 CT scans, using internal five-fold cross-validation and external validations. RECORD with the most effective backbone achieved an average AUC-response of 0.981, AUC-stable of 0.929, and AUC-progression of 0.969 for SOV-based disease status classification, F1-score of 0.887 for new lesion identification, and accuracy of 0.889 for final treatment outcome assessments across all cohorts.

录制管道是由CNN和ViT主干开发的。使用内部五重交叉验证和外部验证，在涉及60个多国中心，206名患者，891次CT扫描的三个纵向队列中评估了其表现。具有最有效骨架的记录实现了基于SOV的疾病状态分类的平均AUC响应为0.981，AUC稳定为0.929，AUC进展为0.969，新病变识别的F1评分为0.887，所有队列的最终治疗结果评估的准确性为0.889。

RECORD’s PFS and response time predictions strongly correlated with clinician’s assessments (P < 0.001). Moreover, RECORD can better stratify high-risk versus low-risk patients for overall survival compared to the human-assessed RECIST results. In conclusion, RECORD demonstrates efficiency and objectivity in analyzing liver lesions for treatment respon.

RECORD的PFS和反应时间预测与临床医生的评估密切相关（P<0.001）。此外，与人类评估的RECIST结果相比，RECORD可以更好地对高风险患者和低风险患者的总体生存率进行分层。总之，记录证明了分析肝脏病变以进行治疗反应的有效性和客观性。

IntroductionObjective response rate and progression-free survival (PFS) are the commonly used endpoints in phase II/III clinical trials of anticancer drugs1,2, which rely heavily on accurate assessment of treatment outcomes. The Response Evaluation Criteria in Solid Tumors (RECIST) guideline (version 1.1) is the currently established, standardized approach for assessing treatment response based on changes in tumor burden3.

引言客观缓解率和无进展生存期（PFS）是抗癌药物II/III期临床试验中常用的终点[1,2]，这在很大程度上依赖于对治疗结果的准确评估。实体瘤反应评估标准（RECIST）指南（1.1版）是目前建立的基于肿瘤负荷变化评估治疗反应的标准化方法3。

RECIST defines four principal tumor response outcomes: complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD), according to the changes in tumor diameter between baseline and follow-up scans. In clinical practice, radiologists must perform several steps to evaluate treatment response: identify lesions on serial imaging, select target lesions, measure changes in tumor burden on serial scans, and make optimal therapeutic decisions.

RECIST根据基线和随访扫描之间肿瘤直径的变化定义了四种主要的肿瘤反应结果：完全缓解（CR），部分缓解（PR），稳定疾病（SD）和进行性疾病（PD）。在临床实践中，放射科医生必须执行几个步骤来评估治疗反应：在连续成像中识别病变，选择目标病变，测量连续扫描中肿瘤负荷的变化，并做出最佳治疗决策。

The procedure is extremely labor-intensive and time-consuming. Furthermore, the subjectivity of image-reading procedures can lead to discrepancies between radiologists. Studies have reported discordance rates ranging from 23% to 46% and 42% between two readers4,5. Inaccurate assessments can lead to wrong medical decisions, thereby compromising treatment efficacy and shortening survival time.Currently, artificial intelligence (AI) models have been widely applied to automatically quantify the tumor burden from CT scans, such as measuring the diameter6,7 and segmenting the 3D mask8,9,10 of a specific lesion, in seconds.

这个过程非常劳动密集且耗时。此外，图像读取程序的主观性可能导致放射科医生之间的差异。研究报道，两名读者之间的不一致率从23%到46%和42%不等4,5。。目前，人工智能（AI）模型已被广泛应用于自动量化CT扫描的肿瘤负荷，例如在几秒钟内测量直径6,7并分割特定病变的3D mask8,9,10。

Consequently, both linear and volumetric measurements can be efficiently obtained, where it was previously arduous for oncologists to manually draw tumor contours on each slice in the clinical practice. Besides, several researchers have investigated the feasibility of vol.

因此，可以有效地获得线性和体积测量，而肿瘤学家在临床实践中手动绘制每个切片上的肿瘤轮廓以前是很困难的。此外，一些研究人员已经调查了vol的可行性。

(1)

(1)

where,

其中，

\({r}_{{PR}}\), \({r}_{{PD}}\) is extrapolated from 1D RECIST to 3D according to our dataset, where we set \({r}_{{PR}}=0.65\), \({r}_{{PD}}=1.44\). The calculation of these two critical values is described in “Validations on the designed mapping function”.

\（笑声）({r}_{{{PR}}\）\({r}_根据我们的数据集，我们将{PD}}从1D RECIST外推到3D\({r}_{{PR}}=0.65 \）\({r}_{{PD}}=1.44 \）。。

\({C}_{i}\) is a constant, where we set \({C}_{1}=15\), \({C}_{2}=5\), \({C}_{3}=0.1\)；

\（笑声）({C}_{i} \）是一个常数，我们在其中设置\({C}_{1} =15 \）\({C}_{2} \({C}_{3} =0.1 \）

\(\sigma\) is the Sigmoid function, which maps \(\left[-\infty ,+\infty \right]\) to [0, 1];

\（\ sigma \）是S形函数，它将\（\ left[-\ infty，+\ infty \ right]\）映射到[0，1]；

\({V}_{1}\) is the baseline SOV of all liver lesions, \({V}_{2}\) is the follow-up SOV of all liver lesions;

\（笑声）({V}_{1} \）是所有肝脏病变的基线SOV\({V}_{2} \）是所有肝脏病变的后续SOV；

\(w\) is the trade-off between relative and absolute volume change, \(w=\sigma \left({C}_{2}\cdot \log \left(\tfrac{\left|{V}_{2}-{V}_{1}\right|}{V}\right)+{C}_{3}\cdot \max \left(\tfrac{{V}_{2}}{V},\,\tfrac{{V}_{1}}{V}\right)\right)\), which works only if both the baseline and follow-up lesions are small.

\（w）是相对体积变化和绝对体积变化之间的权衡，\（w=\ sigma \ left({C}_{2} \cdot\log\left（\tfrac{\left|{V}_{2}-{V}_{1} \右|}{V}\右）+{C}_{3} \cdot\max\left（\tfrac{{V}_{2} }{V}，\，\tfrac{{V}_{1} }{V}\右\右）\），只有在基线和随访病变都很小的情况下才能起作用。

\(V\) is the absolute volume required for progression, where we set \(V=400\). The two items are in line with the RECIST guideline, in which a change in diameter of less than 5 mm would be deemed stable disease;.

\（V \）是进展所需的绝对体积，我们在其中设置\（V=400 \）。这两个项目符合RECIST指南，其中直径变化小于5mm将被认为是稳定的疾病；。

When the SOV is not particularly small (both SOV of baseline and follow are smaller than ~1 cm3), the contribution of the second term can be neglected. The argmax of the three-class probability vector should be equivalent to: a 44% increase indicates progression, and a 35% decrease indicates response..

当SOV不是特别小（基线和后续的SOV都小于〜1cm3）时，第二项的贡献可以忽略不计。三类概率向量的argmax应等于：增加44%表示进展，减少35%表示响应。。

New lesions were evaluated by radiologists. Ground truth for the treatment response evaluation integrated the result of both tumor burden classification and new lesion identification, with the rule shown in Supplementary Table 1.Validations on the designed mapping functionThe mapping function we designed referenced the existing 3D volumetric lesion measurement methods.

放射科医生评估了新的病变。治疗反应评估的基本事实将肿瘤负荷分类和新病变识别的结果与补充表1中所示的规则相结合。对设计的映射函数的验证我们设计的映射函数引用了现有的3D体积病变测量方法。

In the existing literature43,44, there are two different thresholds for progression and response evaluation—one assuming the lesions are spherical (in which case, progression is defined as a 73% or more increase in volume, and response is defined as a 65% or more decrease in volume), and the other assuming the lesions are ellipsoidal (in which case, progression is defined as a 20% or more increase in volume, and response is defined as a 35% or more decrease in volume).

在现有文献43,44中，进展和反应评估有两个不同的阈值，一个假设病变是球形的（在这种情况下，进展定义为体积增加73%或更多，反应定义为体积减少65%或更多），另一个假设病变是椭圆形的（在这种情况下，进展定义为体积增加20%或更多，反应定义为体积减少35%或更多）。

However, in actual data, lesions may not ideally be spherical or ellipsoidal, so using any one of the two thresholds could introduce potential biases. Consequently, we aimed to establish the critical volume change thresholds for progression and response by calculating the average volume ratio change for a single lesion when its diameter increased by 20% or decreased by 30% in our dataset.

然而，在实际数据中，病变可能不理想地是球形或椭圆形，因此使用两个阈值中的任何一个都可能引入潜在的偏差。因此，我们的目标是通过计算单个病变的平均体积比变化来确定进展和反应的临界体积变化阈值，当其直径在我们的数据集中增加20%或减少30%时。

Cohort A was conducted with manual diameter measurements and was therefore used for mapping function modeling and validation. Specifically, we selected the single lesion in Cohort A1 that had a diameter reduction of 25–35%, and calculated the average volume change ratio of 0.6508 ± 0.1051. Similarly, lesions with a diameter increase of 15–25% yielded an average volume change ratio of 1.4403 ± 0.2282.

。具体而言，我们在队列A1中选择了直径减小25-35%的单个病变，并计算出平均体积变化率为0.6508±0.1051。同样，直径增加15-25%的病变的平均体积变化率为1.4403±0.2282。

Validation results in Cohort A2 showed that the average volume change ratio was 0.6398 ± 0.0531, and 1.4639 ± 0.1379 for response and progression, respectivel.

队列A2的验证结果显示，反应和进展的平均体积变化率分别为0.6398±0.0531和1.4639±0.1379。

(2)

(2)

$${L}_{{seg}}={L}_{{seg}{{\_}}{baseline}}+{L}_{{seg}{{\_}}{followup}}$$

${L}__其他组织者{L}_｛获取｝｛｛\｝｛基线｝+{L}_｛获取｝｛｛\｝｛跟进｝$

(3)

(3)

$${L}_{{progression}}=-{y\,\cdot\,ln}\hat{y},\hat{y}=\frac{{e}^{-{x}_{i}}}{\mathop{\sum }\nolimits_{k=1}^{3}{e}^{-{x}_{k}}}i=1,2,3$$

（笑声）$${L}_^{-{x}_{i} }}{\mathop{\sum}\nolimits\uk=1}^{3}{e}^{-{x}_{k} }}i=1,2,3$$

(4)

(4)

where \(w\) is the trade-off between two tasks. Based on our experiments, we recommend setting \(w\) between 0.4 and 0.7, and found that \(w=0.55\) achieve the best performance. \({L}_{{seg}}\) represents the segmentation loss for both the baseline and follow-up images. \({L}_{{progression}}\) represents the progression loss, which depicts the classification accuracy of response evaluations.

。根据我们的实验，我们建议将（w）设置在0.4到0.7之间，并发现（w=0.55）达到最佳性能\（笑声）({L}_{{seg}}表示基线和后续图像的分割损失\（笑声）({L}_{{进展}}表示进展损失，它描述了响应评估的分类准确性。

It is crucial to highlight that the ground truth label y is designed to be continuous in the model optimization, without employing argmax that is typically utilized in traditional classification problems. This is because there exists an order among response, stable, and progressive, thus regarding this as an ordinal regression problem, rather than a traditional classification problem, can give more tolerance to corner cases.The SOV-based treatment response classification model was implemented using MONAI 0.9.1, a PyTorch-based repository.

至关重要的是要强调，地面真值标签y在模型优化中被设计为连续的，而不使用传统分类问题中通常使用的argmax。这是因为响应，稳定和渐进之间存在顺序，因此将其视为序数回归问题，而不是传统的分类问题，可以对角落案例提供更多的容忍度。基于SOV的治疗反应分类模型是使用基于PyTorch的存储库MONAI 0.9.1实现的。

To handle the computation load of 3D image pairings, we adopted data parallel, used sliding window strategies to distribute ROI patches computation to four GPUs.Explanations on the design of ordinal regression-based loss \({{\boldsymbol{L}}}_{{\boldsymbol{progression}}}\).

为了处理3D图像配对的计算负载，我们采用数据并行，使用滑动窗口策略将ROI补丁计算分配到四个GPU。关于基于序数回归的损失设计的解释“（{{\ boldsymbol{L}}}}}{\ boldsymbol{progression}}}”。

The loss function of response outcome classification is based on ordinal regression, which is modified from the traditional categorical classification to in line with the continuous disease progression process. Traditional cross-entropy loss treats the ground truth label as three one-hot classes, while our modified loss function converts the discrete three-class label into a probability vector.

响应-结果分类的损失函数基于序数回归，该回归从传统的分类分类修改为符合连续的疾病进展过程。传统的交叉熵损失将地面真值标签视为三个一个热类，而我们改进的损失函数将离散的三类标签转换为概率向量。

The nature of treatment outcome is a continuous quantification of change ratio of tumor burden, rather than a discrete value determined by several cut-offs. Specifically, thresholds between disease statuses are manually set to get the results from the baseline-follow-up SOV change ratio, with a given percentage of SOV decrease denotes PR (in our scenario is 35%), and another given percentage increase in SOV denotes PD (in our scenario is 44%).

治疗结果的性质是肿瘤负荷变化率的连续量化，而不是由几个临界值确定的离散值。具体而言，手动设置疾病状态之间的阈值，以获得基线随访SOV变化率的结果，给定的SOV下降百分比表示PR（在我们的场景中为35%），另一个给定的SOV增加百分比表示PD（在我们的场景中为44%）。

One-hot discrete labels can have information loss. For instance, clinicians would consider 43% and 45% increase in SOV have little difference in disease progression status. However, if we adopt categorical classification during the optimization, the former would be one-hot encoded with (0, 1, 0) (SD), while the latter would be (0, 0, 1) (PD), leading to totally different ground truth response outcomes.

一个热离散标签可能会丢失信息。例如，临床医生会认为SOV增加43%和45%在疾病进展状态上几乎没有差异。然而，如果我们在优化过程中采用分类分类，前者将是一个用（0,1,0）（SD）热编码的分类，而后者将是（0,0,1）（PD），导致完全不同的地面真相响应结果。

The simplest way to avoid information loss is to directly predict the continuous tumor burden change ratio. Whereas, it is also inappropriate to treat the problem as a traditional regression task that employs a standard mean squared error (MSE) loss, since the tumor burden change ratio exhibits boundary effects.

避免信息丢失的最简单方法是直接预测连续的肿瘤负荷变化率。然而，将该问题视为采用标准均方误差（MSE）损失的传统回归任务也是不合适的，因为肿瘤负荷变化率表现出边界效应。

For example, extreme change ratios such as 20 versus 30 times greater may both be indicative of disease progression, yet they can cause large loss during model optimization when using MSE as the loss function. Conversely, small change.

例如，极端变化率（例如20倍对30倍）可能都表明疾病进展，但当使用MSE作为损失函数时，它们可能在模型优化过程中造成巨大损失。相反，变化很小。

Data availability

数据可用性

The datasets cannot be made publicly available due to general data protection regulations and institutional guidelines. Data can be retrieved upon reasonable request.

由于一般数据保护法规和机构指南，这些数据集无法公开提供。可根据合理要求检索数据。

Code availability

代码可用性

The study source code can be found at https://github.com/EstelleXIA/RECORD/.

研究源代码可以在https://github.com/EstelleXIA/RECORD/.

ReferencesRitchie, G. et al. Defining the most appropriate primary end point in phase 2 trials of immune checkpoint inhibitors for advanced solid cancers. JAMA Oncol. 4, 522 (2018).Article

参考文献Ritchie，G。等人在晚期实体癌免疫检查点抑制剂的2期试验中定义了最合适的主要终点。JAMA Oncol。4522（2018）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Teuwen, L.-A. M. N., Young, J. A., Bourlon, M. T., Segelov, E. & Prenen, H. Endpoints reported in phase 3 randomized clinical trials at ASCO 2022. J. Clin. Oncol. 41, 1570 (2023).Article

Teuwen，L.-A.M.N.，Young，J.A.，Bourlon，M.T.，Segelov，E.＆Prenen，H.终点在ASCO 2022的3期随机临床试验中报道。J、 临床。。。文章

Google Scholar

谷歌学者

Eisenhauer, E. et al. New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1). Eur. J. Cancer 45, 228–247 (2009).Article

Eisenhauer，E.等人，《实体瘤新的反应评估标准：修订的RECIST指南（1.1版）》。《欧洲癌症杂志》45228-247（2009）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Gonen, CohenL. & Ford, M. R. Monitoring reader metrics in blinded independent central review of oncology studies. J. Clin. Trials 05, 4 (2015).

科恩·戈恩和Ford，M.R。在肿瘤学研究的盲法独立中央审查中监测读者指标。J、 临床。试验05,4（2015）。

Google Scholar

谷歌学者

Ford, R., Neal, M., Moskowitz, S. & Fraunberger, J. Adjudication rates between readers in blinded independent central review of oncology studies. J. Clin. Trials 06, 5 (2016).

Ford，R.，Neal，M.，Moskowitz，S。和Fraunberger，J。在肿瘤学研究的盲法独立中央审查中，读者之间的裁决率。J、 临床。试验06,5（2016）。

Google Scholar

谷歌学者

Rafael-Palou, X. et al. Re-Identification and growth detection of pulmonary nodules without image registration using 3D Siamese neural networks. Med. Image Anal. 67, 101823 (2021).Article

Rafael Palou，X。等人。使用3D暹罗神经网络在没有图像配准的情况下重新识别和生长检测肺结节。医学图像肛门。67101823（2021）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Cai, J. et al. Deep volumetric universal lesion detection using Light-Weight pseudo 3D convolution and surface point regression. Lect. Notes Comput. Sci. 12264, 3–13 (2020).Article

Cai，J。等人。使用轻质伪3D卷积和表面点回归的深度体积通用病变检测。选择。注释计算机。科学。12264，3-13（2020）。文章

Google Scholar

谷歌学者

Primakov, S. et al. Automated detection and segmentation of non-small cell lung cancer computed tomography images. Nat. Commun. 13, 3423 (2022).Article

Primakov，S.等人。非小细胞肺癌计算机断层扫描图像的自动检测和分割。国家公社。133423（2022）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Antonelli, M. et al. The medical segmentation decathlon. Nat. Commun. 13, 4128 (2022).Article

Antonelli，M.等人，《医学分段十项全能》。国家公社。134128（2022）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Lv, P., Wang, J., Zhang, X. & Shi, C. Deep supervision and atrous inception-based U-Net combining CRF for automatic liver segmentation from CT. Sci. Rep. 12, 16995 (2022).Article

Lv，P.，Wang，J.，Zhang，X。＆Shi，C。深度监督和基于Atrus inception的U-Net结合CRF用于CT自动肝脏分割。Sci。第1216995页（2022年）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Arshad, M. et al. Assessment of tumor burden and response by RECIST vs. volume change in HPV+ oropharyngeal cancer—an exploratory analysis of prospective trials. Int. J. Radiat. Oncol., Biol., Phys. 114, S113–S114 (2022).Article

Arshad，M.等人。RECIST与HPV+口咽癌体积变化对肿瘤负荷和反应的评估-前瞻性试验的探索性分析。内景J.辐射。Oncol。，生物学，物理。114，S113–S114（2022）。文章

Google Scholar

谷歌学者

Minault, Q., Barthélémy, P., Leyendecker, P., Mielcarek, M. & Roy, C. 740P Response assessment in advanced renal cell carcinoma (mRCC) patients (pts) treated by Nivolumab (N)+ Ipilimumab (I): CT volumetric measurement versus RECIST 1.1 response criteria. Ann. Oncol. 31, S576 (2020).An, Y.

Minault，Q.，Barthélémy，P.，Leyendecker，P.，Mielcarek，M。＆Roy，C。740P在Nivolumab（N）+Ipilimumab（I）治疗的晚期肾细胞癌（mRCC）患者（pts）中的反应评估：CT体积测量与RECIST 1.1反应标准。Ann。Oncol。31，S576（2020）。。

Y., Kim, S. H., Kang, B. J., Lee, A. W. & Song, B. J. MRI volume measurements compared with the RECIST 1.1 for evaluating the response to neoadjuvant chemotherapy for mass-type lesions. Breast Cancer 21, 316–324 (2012).Article .

Y、 ，Kim，S.H.，Kang，B.J.，Lee，A.W。＆Song，B.J。MRI体积测量与RECIST 1.1相比，用于评估肿块型病变对新辅助化疗的反应。乳腺癌21316-324（2012）。文章。

PubMed

PubMed

Google Scholar

谷歌学者

Kalbande, P. B., Aher, P., Kale, P. & Datta, N. R. Comparative evaluation of the sum of longest diameter measurements as per RECIST 1.1 vs. CECT based volumetric estimation for response assessment in locally advanced head and neck cancer. Int. J. Radiat. Oncol., Biol., Phys. 114, e293 (2022).Article .

Kalbande，P.B.，Aher，P.，Kale，P。＆Datta，N.R。根据RECIST 1.1与基于CECT的体积估计的最长直径测量总和的比较评估，用于局部晚期头颈癌的反应评估。内景J.辐射。Oncol。，生物学，物理。114，e293（2022）。文章。

Google Scholar

谷歌学者

Fenerty, K. E. et al. Predicting clinical outcomes in chordoma patients receiving immunotherapy: a comparison between volumetric segmentation and RECIST. BMC Cancer 16, 672 (2016).Article

Fenerty，K.E.等人。预测接受免疫治疗的脊索瘤患者的临床结果：体积分割和RECIST之间的比较。。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Yu, S. C. H., Yeung, D. T. K. & So, N. M. C. Imaging features of hepatocellular carcinoma. Clin. Radiol. 59, 145–156 (2004).Article

Yu，S.C.H.，Yeung，D.T.K.＆So，N.M.C.肝细胞癌的影像学特征。临床。放射性。59145-156（2004）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Reynolds, A. R. et al. Infiltrative hepatocellular carcinoma: what radiologists need to know. Radiographics 35, 371–386 (2015).Article

Reynolds，A.R.等人，《浸润性肝细胞癌：放射科医生需要知道的》。射线照相35371-386（2015）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Tacher, V. et al. Comparison of existing response criteria in patients with hepatocellular carcinoma treated with transarterial chemoembolization using a 3D quantitative approach. Radiology 278, 275–284 (2016).Article

Tacher，V。等人。使用3D定量方法比较经动脉化疗栓塞治疗的肝细胞癌患者的现有反应标准。放射学278275-284（2016）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Lee, I. S., Choi, S. J., Seo, C. R. & Kim, J. S. Comparison of the response evaluation criteria in solid tumors with volumetric measurement for evaluation of response and overall survival with liver metastases from colorectal cancer. J. Korean Soc. Radiol. 80, 906 (2019).Article

Lee，I.S.，Choi，S.J.，Seo，C.R。＆Kim，J.S。实体瘤疗效评估标准与体积测量的比较，用于评估结直肠癌肝转移的疗效和总生存率。J、 韩国社会放射学。80906（2019）。文章

Google Scholar

谷歌学者

Doemel, L. A. et al. Reliable prediction of survival in advanced-stage hepatocellular carcinoma treated with sorafenib: comparing 1D and 3D quantitative tumor response criteria on MRI. Eur. Radiol. 31, 2737–2746 (2020).Article

Doemel，L.A.等人，《索拉非尼治疗晚期肝细胞癌生存率的可靠预测：MRI上1D和3D定量肿瘤反应标准的比较》，欧洲放射学杂志。312737-2746（2020）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Xue, Z., Shen, D. & Davatzikos, C. CLASSIC: consistent longitudinal alignment and segmentation for serial image computing. NeuroImage 30, 388–399 (2006).Article

Xue，Z.，Shen，D。和Davatzikos，C。经典：连续图像计算的一致纵向对齐和分割。神经影像30388-399（2006）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Wei, J. et al. Consistent segmentation of longitudinal brain MR images with spatio-temporal constrained networks. Med. Image Comput. Comput. Assist. Intervention 12901, 89–98 (2021).

Wei，J.等人。使用时空约束网络对纵向脑MR图像进行一致分割。医学图像计算机。计算机。协助。干预12901,89-98（2021）。

Google Scholar

谷歌学者

Ansari, M. Y. et al. Practical utility of liver segmentation methods in clinical surgeries and interventions. BMC Med. Imaging 22, 97 (2022).Article

Ansari，M.Y.等人。肝脏分割方法在临床手术和干预中的实际应用。BMC医学影像22,97（2022）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Anderson, B. M. et al. Automated segmentation of colorectal liver metastasis and liver ablation on contrast-enhanced CT images. Front. Oncol. 12, 886517 (2022).Article

Anderson，B.M.等人。对比增强CT图像上结直肠肝转移和肝脏消融的自动分割。正面。。12886517（2022年）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Vivanti, R., Szeskin, A., Lev‐Cohain, N., Sosna, J. & Joskowicz, L. Automatic detection of new tumors and tumor burden evaluation in longitudinal liver CT scan studies. Int. J. Comput. Assist. Radiol. Surg. 12, 1945–1957 (2017).Article

Vivanti，R.，Szeskin，A.，Lev-Cohain，N.，Sosna，J。＆Joskowicz，L。纵向肝脏CT扫描研究中新肿瘤的自动检测和肿瘤负荷评估。国际J.计算机。协助。放射性。外科杂志121945-1957（2017）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Isensee, F., Jaeger, P. F., Kohl, S., Petersen, J. & Maier‐Hein, K. H. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods 18, 203–211 (2020).Article

Isensee，F.，Jaeger，P.F.，Kohl，S.，Petersen，J.＆Maier-Hein，K.H.nnU Net：一种基于深度学习的生物医学图像分割的自配置方法。自然方法18203-211（2020）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Hatamizadeh, A. et al. SWIN UNETR: SWIN transformers for semantic segmentation of brain tumors in MRI images. Lect. Notes Comput. Sci. 12962, 272–284 (2022).Article

Hatamizadeh，A。et al。SWIN UNETR：用于MRI图像中脑肿瘤语义分割的SWIN变压器。选择。注释计算机。科学。12962272-284（2022）。文章

Google Scholar

谷歌学者

Arbour, K. C. et al. Deep learning to estimate RECIST in patients with NSCLC treated with PD-1 blockade. Cancer Discov. 11, 59–67 (2021).Article

Arbour，K.C.等人。深入学习评估PD-1阻断治疗的NSCLC患者的RECIST。癌症发现。11，59-67（2021）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Tang, Y. et al. Accurate and robust lesion RECIST diameter prediction and segmentation with transformers. Lect. Notes Comput. Sci. 13434, 535–544 (2022).Article

Tang，Y.等人。使用变压器进行准确而稳健的病变RECIST直径预测和分割。选择。注释计算机。科学。13434535-544（2022）。文章

Google Scholar

谷歌学者

Tang, Y. et al. Lesion segmentation and RECIST diameter prediction via click-driven attention and dual-path connection. Lect. Notes Comput. Sci. 12902, 341–351 (2021).Article

Tang，Y.等人。通过点击驱动的注意力和双路径连接进行病变分割和RECIST直径预测。选择。注释计算机。科学。12902341-351（2021）。文章

Google Scholar

谷歌学者

Vivanti, R., Joskowicz, L., Lev‐Cohain, N., Ephrat, A. & Sosna, J. Patient-specific and global convolutional neural networks for robust automatic liver tumor delineation in follow-up CT studies. Med. Biol. Eng. Comput. 56, 1699–1713 (2018).Article

Vivanti，R.，Joskowicz，L.，Lev-Cohain，N.，Ephrat，A。＆Sosna，J。患者特异性和全局卷积神经网络，用于在后续CT研究中进行稳健的自动肝肿瘤描绘。医学生物学。工程计算。561699-1713（2018）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Fang, J. et al. Siamese encoder-based spatial-temporal mixer for growth trend prediction of lung nodules on CT scans. Lect. Notes Comput. Sci. 13431, 484–494 (2022).Article

。选择。注释计算机。科学。13431484-494（2022）。文章

Google Scholar

谷歌学者

Cai, J. et al. Deep lesion tracker: monitoring lesions in 4D longitudinal imaging studies. In IEEE Conference on Computer Vision and Pattern Recognition, 15154–15164 (2021).Yang, Y., Yang, J., Ye, Y., Xia, T. & Lu, S. Development and validation of a deep learning model to assess tumor progression to immunotherapy.

蔡，J。等。深部病变跟踪器：在4D纵向成像研究中监测病变。在IEEE计算机视觉和模式识别会议上，15154–15164（2021）。Yang，Y.，Yang，J.，Ye，Y.，Xia，T。＆Lu，S。开发和验证深度学习模型以评估肿瘤进展至免疫疗法。

J. Clin. Oncol. 37, e20601 (2019).Article .

J.克林顿。肿瘤。37，e20601（2019）。文章联盟。

Google Scholar

谷歌学者

Moreau, N. et al. Automatic segmentation of metastatic breast cancer lesions on 18F-FDG PET/CT longitudinal acquisitions for treatment response assessment. Cancers 14, 101 (2021).Article

Moreau，N.等人。18F-FDG PET/CT纵向采集中转移性乳腺癌病变的自动分割，用于治疗反应评估。癌症14101（2021）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Zhou, J., Xia, Y., Xun, X. & Yu, Z. Deep learning-based detect-then-track pipeline for treatment outcome assessments in immunotherapy-treated liver cancer. J. Imaging Inform. Med. https://doi.org/10.1007/s10278-024-01132-8 (2024).Bruix, J., Da Fonseca, L. G. & Reig, M. Insights into the success and failure of systemic therapy for hepatocellular carcinoma.

Zhou，J.，Xia，Y.，Xun，X。＆Yu，Z。基于深度学习的检测然后跟踪管道，用于免疫疗法治疗的肝癌的治疗结果评估。J、 成像通知。医学。https://doi.org/10.1007/s10278-024-01132-8（2024年）。Bruix，J.，Da Fonseca，L.G。＆Reig，M。洞察肝细胞癌全身治疗的成功与失败。

Nat. Rev. Gastroenterol. Hepatol. 16, 617–630 (2019).Article .

国家胃肠病学杂志。Hepatol。16, 617–630 (2019).第条。

PubMed

PubMed

Google Scholar

谷歌学者

Bruix, J. Endpoints in clinical trials for liver cancer and their value in evidence-based clinical decision making: an unresolved Gordian knot. J. Hepatol. 74, 1483–1488 (2021).Article

Bruix，J。肝癌临床试验的终点及其在循证临床决策中的价值：尚未解决的棘手问题。J、 肝病。741483-1488（2021）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Lencioni, R. & Llovet, J. Modified RECIST (MRECIST) assessment for hepatocellular carcinoma. Semin. Liver Dis. 30, 052–060 (2010).Article

Lencioni，R。＆Llovet，J。改良RECIST（MRECIST）评估肝细胞癌。塞米。肝脏疾病。3052-060（2010）。文章

Google Scholar

谷歌学者

Lowekamp, B., Chen, D. T., Ibáñez, L. & Blezek, D. J. The design of SimpleITK. Front. Neuroinform. 7, 45 (2013).Article

Lowekamp，B.，Chen，D.T.，Ibáñez，L.&Blezek，D.J.SimpleITK的设计。前面。神经信息学家。7, 45 (2013).文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Shrout, P. E. & Fleiss, J. L. Intraclass correlations: uses in assessing rater reliability. Psychol. Bull. 86, 420–428 (1979).Article

Shrout，P.E。＆Fleiss，J.L。组内相关性：用于评估评估者的可靠性。心理学。公牛。86420-428（1979）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Krippendorff, K. Content Analysis. An Introduction to Its Methodology (3rd edn) (Sage Publications, 2013).Schiavon, G. et al. Tumor volume as an alternative response measurement for imatinib treated GIST patients. PLoS ONE 7, e48372 (2012).Article

Krippendorff，K。含量分析。其方法简介（第三版）（Sage Publications，2013）。Schiavon，G。等人。肿瘤体积作为伊马替尼治疗的GIST患者的替代反应测量。PLoS ONE 7，e48372（2012）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Levine, Z. H. et al. RECIST versus volume measurement in medical CT using ellipsoids of known size. Opt. Express 18, 8151 (2010).Article

Levine，Z.H.等人。RECIST与使用已知大小椭球体的医学CT中的体积测量。选择。快报188151（2010）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Avants, B. B., Tustison, N. & Song, G. Advanced normalization tools (ANTS). Insights 2, 1–35 (2009).

Avants，B.B.，Tustison，N。＆Song，G。高级标准化工具（ANTS）。。

Google Scholar

谷歌学者

Pedersen, A. & Pérez de Frutos, J. andreped/livermask: v1.4.1. Zenodo. https://doi.org/10.5281/zenodo.7574587 (2023).Isensee, F., Jaeger, P. F., Kohl, S., Petersen, J. & Maier‐Hein, K. H. Pretrained models for 3D semantic image segmentation with nnU-Net (Version 1). Zenodo. https://doi.org/10.5281/zenodo.3734294 (2020).Hu, J., Shen, L.

Pedersen，A.＆Pérez de Frutos，J。andreped/livermask:v1.4.1。泽诺多。https://doi.org/10.5281/zenodo.7574587（2023年）。Isensee，F.，Jaeger，P.F.，Kohl，S.，Petersen，J.＆Maier-Hein，K.H。使用nnU-Net（版本1）进行3D语义图像分割的预训练模型。泽诺多。https://doi.org/10.5281/zenodo.3734294（2020年）。胡，J.，沈，L。

& Sun, G. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 7132–7141 (2018).Ronneberger, O., Fischer, P. & Brox, T. U-Net: convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Vol.

&Sun，G。挤压和激发网络。在IEEE计算机视觉和模式识别会议论文集，7132-7141（2018）。Ronneberger，O.，Fischer，P。＆Brox，T.U-Net：用于生物医学图像分割的卷积网络。医学图像计算和计算机辅助干预-MICCAI 2015，Vol。

9351, 234–241 (2015).Wilson, E. B. Probable inference, the Law of succession, and statistical inference. J. Am. Stat. Assoc. 22, 209–212 (1927).Article .

9351234-241（2015）。Wilson，E.B。或然推理，继承定律和统计推理。J、 美国统计协会第22209-212号（1927年）。文章。

Google Scholar

谷歌学者

DeLong, E. R., DeLong, D. M. & Clarke‐Pearson, D. L. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 44, 837 (1988).Article

DeLong，E.R.，DeLong，D.M。＆Clarke-Pearson，D.L。比较两个或多个相关接收器工作特性曲线下的面积：非参数方法。生物特征44837（1988）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Fagerland, M. W., Lydersen, S. & Laake, P. The McNemar test for binary matched-pairs data: mid-p and asymptotic are better than exact conditional. BMC Med. Res. Methodol. 13, 91 (2013).Article

Fagerland，M.W.，Lydersen，S。＆Laake，P。二元匹配对数据的McNemar检验：mid-P和渐近优于精确条件。BMC医学研究方法。13，91（2013）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Fan, Y. & Yin, G. Concordance index: Surrogacy of progression-free survival for overall survival. Contemp. Clin. Trials 104, 106353 (2021).Article

Fan，Y。＆Yin，G。一致性指数：无进展生存期替代总生存期。康坦普。临床。试验104106353（2021）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Eden, S. K., Li, C. & Shepherd, B. E. Nonparametric estimation of Spearman’s rank correlation with bivariate survival data. Biometrics 78, 421–434 (2021).Article

Eden，S.K.，Li，C。＆Shepherd，B.E。Spearman等级相关性与双变量生存数据的非参数估计。生物特征78421-434（2021）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Halabi, S. et al. Progression-free survival as a predictor of overall survival in men with castrate-resistant prostate cancer. J. Clin. Oncol. 27, 2766–2771 (2009).Article

Halabi，S.等人。无进展生存期作为去势抵抗性前列腺癌男性总生存期的预测指标。J、 临床。。272766-2771（2009）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Download referencesAcknowledgementsThis study was supported by National Natural Science Foundation of China (ID: 12171318), Shanghai Science and Technology Commission (ID: 21ZR1436300), Shanghai Jiao Tong University STAR Grant (ID: 20190102), Medical Engineering Cross Fund of Shanghai Jiao Tong University (ID: YG2023ZD21).

下载参考文献致谢本研究得到了国家自然科学基金（编号：12171318），上海市科学技术委员会（编号：21ZR1436300），上海交通大学明星基金（编号：20190102），上海交通大学医学工程交叉基金（编号：YG2023ZD21）的支持。

The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.Author informationAuthor notesThese authors contributed equally: Yujia Xia, Jie Zhou.Authors and AffiliationsDepartment of Bioinformatics and Biostatistics, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, 200240, ChinaYujia Xia, Ting Wei, Ruitian Gao, Yufei Zhang & Zhangsheng YuSJTU-Yale Joint Center for Biostatistics and Data Science, Shanghai Jiao Tong University, Shanghai, 200240, ChinaYujia Xia, Jie Zhou, Luke Johnston, Ting Wei, Ruitian Gao, Yufei Zhang & Zhangsheng YuDepartment of Statistics, School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, 200240, ChinaJie Zhou, Luke Johnston & Zhangsheng YuStatistics in Global Statistics and Data Science, Beigene, Shanghai, 200040, ChinaXiaolei XunPiHealth USA, 55 Cambridge Parkway, Cambridge, MA, 02142, USABobby Reddy, Chao Liu, Geoffrey Kim & Jin ZhangDepartment of Transplantation, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, 200092, ChinaShuai ZhaoClinical Research Institute, Shanghai Jiao Tong University School of Medicine, Shanghai, 200025, ChinaZhangsheng YuAuthorsYujia XiaView author publicationsYou can also search for this author in.

资助者在研究设计，数据收集，数据分析和解释或撰写本手稿方面没有发挥任何作用。作者信息作者注意到这些作者做出了同样的贡献：夏玉嘉，周杰。作者和所属单位上海交通大学生命科学与生物技术学院生物信息学与生物统计学系，上海，200240，中国夏玉嘉，丁伟，高瑞田，张玉飞和张生，上海交通大学耶鲁生物统计学与数据科学联合中心，上海，200240，中国夏玉嘉，周杰，卢克·约翰斯顿，丁伟，高瑞田，张玉飞和张生，上海交通大学数理科学学院统计系，上海，200240，周杰，卢克·约翰斯顿和张生，全球统计与数据科学统计，贝根，上海，200040上海交通大学医学院附属新华医院移植科，上海，200092，上海交通大学医学院临床研究所，上海，200025，中国张生YuAuthorsYujia XiaView作者出版物您也可以在中搜索此作者。

PubMed Google ScholarJie ZhouView author publicationsYou can also search for this author in

PubMed Google ScholarJie ZhouView作者出版物您也可以在

PubMed Google ScholarXiaolei XunView author publicationsYou can also search for this author in

PubMed Google ScholarLuke JohnstonView author publicationsYou can also search for this author in

PubMed Google ScholarLuke JohnstonView作者出版物您也可以在

PubMed Google ScholarTing WeiView author publicationsYou can also search for this author in

PubMed Google ScholarTing WeiView作者出版物您也可以在

PubMed Google ScholarRuitian GaoView author publicationsYou can also search for this author in

PubMed Google ScholarRuitian GaoView作者出版物您也可以在

PubMed Google ScholarYufei ZhangView author publicationsYou can also search for this author in

PubMed Google ScholarYufei ZhangView作者出版物您也可以在

PubMed Google ScholarBobby ReddyView author publicationsYou can also search for this author in

PubMed Google ScholarBobby ReddyView作者出版物您也可以在

PubMed Google ScholarChao LiuView author publicationsYou can also search for this author in

PubMed Google ScholarChao LiuView作者出版物您也可以在

PubMed Google ScholarGeoffrey KimView author publicationsYou can also search for this author in

PubMed Google ScholarGeoffrey KimView作者出版物您也可以在

PubMed Google ScholarJin ZhangView author publicationsYou can also search for this author in

PubMed Google ScholarJin ZhangView作者出版物您也可以在

PubMed Google ScholarShuai ZhaoView author publicationsYou can also search for this author in

PubMed Google ScholarShuai ZhaoView作者出版物您也可以在

PubMed Google ScholarZhangsheng YuView author publicationsYou can also search for this author in

PubMed谷歌学者张生YuView作者出版物您也可以在

PubMed Google ScholarContributionsConceptualization: Z.S.Y. and JinZ. Methodology: Y.J.X. and JieZ. Data collection: X.L.X. and S.Z. Investigation: Y.J.X., JieZ., and X.L.X. Visualization: Y.J.X. and R.T.G. Supervision: T.W., B.R., G.K., and C.L. Writing—original draft: Y.J.X.

PubMed谷歌学术贡献概念：Z.S.Y.和JinZ。方法：Y.J.X.和JieZ。数据收集：X.L.X.和S.Z.调查：Y.J.X.，JieZ。，和X.L.X.可视化：Y.J.X.和R.T.G.监督：T.W.，B.R.，G.K。和C.L.撰写原稿：Y.J.X。

Writing—review & editing: JieZ., S.Z., L.J., Z.S.Y., and Y.F.Z.Corresponding authorsCorrespondence to.

写作评论与编辑：JieZ。，S、 。

Jin Zhang, Shuai Zhao or Zhangsheng Yu.Ethics declarations

张晋、赵帅或张生宇。道德宣言

Competing interests

相互竞争的利益

Z.S.Y., Y.J.X., JieZ., JinZ., X.L.X., and G.K. are filing a patent for the deep learning model in this study. All other authors declare they have no competing interests.

Z、 S.Y.，Y.J.X.，杰兹。，不幸的是。，十、 在这项研究中，L.X.和G.K.正在为深度学习模型申请专利。所有其他作者都声明他们没有利益冲突。

Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Supplementary informationSupplementary InformationRights and permissions

。补充信息补充信息权限

Open Access 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.

开放获取本文是根据知识共享署名非商业性NoDerivatives 4.0国际许可证授权的，该许可证允许以任何媒介或格式进行任何非商业性使用，共享，分发和复制，只要您对原始作者和来源给予适当的信任，提供知识共享许可证的链接，并指出您是否修改了许可材料。

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，您无权共享源自本文或其部分的改编材料。本文中的图像或其他第三方材料包含在文章的知识共享许可证中，除非该材料的信用额度中另有说明。如果材料未包含在文章的知识共享许可中，并且您的预期用途不受法律法规的许可或超出许可用途，则您需要直接获得版权所有者的许可。

To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/..

要查看此许可证的副本，请访问http://creativecommons.org/licenses/by-nc-nd/4.0/..

Reprints and permissionsAbout this articleCite this articleXia, Y., Zhou, J., Xun, X. et al. Deep learning for oncologic treatment outcomes and endpoints evaluation from CT scans in liver cancer.

转载和许可本文引用本文Xia，Y.，Zhou，J.，Xun，X。等人。肝癌CT扫描对肿瘤治疗结果和终点评估的深度学习。

npj Precis. Onc. 8, 263 (2024). https://doi.org/10.1038/s41698-024-00754-zDownload citationReceived: 16 April 2024Accepted: 04 November 2024Published: 17 November 2024DOI: https://doi.org/10.1038/s41698-024-00754-zShare this articleAnyone you share the following link with will be able to read this content:Get shareable linkSorry, a shareable link is not currently available for this article.Copy to clipboard.

npj精度。Onc。8263（2024）。https://doi.org/10.1038/s41698-024-00754-zDownload引文收到日期：2024年4月16日接受日期：2024年11月4日发布日期：2024年11月17日OI：https://doi.org/10.1038/s41698-024-00754-zShare本文与您共享以下链接的任何人都可以阅读此内容：获取可共享链接对不起，本文目前没有可共享的链接。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