AbstractThe rise of object detection models has brought new breakthroughs to the development of clinical decision support systems. However, in the field of gastrointestinal polyp detection, there are still challenges such as uncertainty in polyp identification and inadequate coping with polyp scale variations.

目标检测模型的兴起为临床决策支持系统的发展带来了新的突破。然而，在胃肠道息肉检测领域，仍然存在挑战，例如息肉识别的不确定性和应对息肉规模变化的不足。

To address these challenges, this paper proposes a novel gastrointestinal polyp object detection model. The model can automatically identify polyp regions in gastrointestinal images and accurately label them. In terms of design, the model integrates multi-channel information to enhance the ability and robustness of channel feature expression, thus better coping with the complexity of polyp structures.

为了应对这些挑战，本文提出了一种新的胃肠息肉目标检测模型。该模型可以自动识别胃肠图像中的息肉区域并准确标记它们。在设计方面，该模型整合了多通道信息，增强了通道特征表达的能力和鲁棒性，从而更好地应对息肉结构的复杂性。

At the same time, a hierarchical structure is constructed in the model to enhance the model’s adaptability to multi-scale targets, effectively addressing the problem of large-scale variations in polyps. Furthermore, a channel attention mechanism is designed in the model to improve the accuracy of target positioning and reduce uncertainty in diagnosis.

同时，在模型中构建了层次结构，增强了模型对多尺度目标的适应性，有效地解决了息肉大规模变化的问题。此外，该模型还设计了一种通道注意机制，以提高目标定位的准确性，减少诊断的不确定性。

By integrating these strategies, the proposed gastrointestinal polyp object detection model can achieve accurate polyp detection, providing clinicians with reliable and valuable references. Experimental results show that the model exhibits superior performance in gastrointestinal polyp detection, which helps improve the diagnostic level of digestive system diseases and provides useful references for related research fields..

。实验结果表明，该模型在胃肠道息肉检测中表现出优越的性能，有助于提高消化系统疾病的诊断水平，为相关研究领域提供有用的参考。。

IntroductionWith the continuous changes in human dietary habits, gastrointestinal polyps have become an increasingly common digestive system disease, posing a significant threat to the lives and health of numerous patients1,2. While most polyps are benign, the potential risk of malignant transformation should not be overlooked, making timely detection and treatment crucial3,4.

引言随着人类饮食习惯的不断变化，胃肠道息肉已成为越来越常见的消化系统疾病，对众多患者的生命和健康构成重大威胁1,2。虽然大多数息肉是良性的，但不应忽视恶变的潜在风险，因此及时发现和治疗至关重要3,4。

Although conventional endoscopic examination methods are commonly used, their reliance on the experience and skills of medical professionals, coupled with the possibility of misdiagnosis, no longer suffice to meet the growing diagnostic demands. Therefore, developing an accurate, rapid, and user-friendly means of detecting gastrointestinal polyps has become an urgent task in current medical research5,6.In recent years, deep learning technologies have achieved remarkable success in the field of computer vision, opening up new possibilities for medical image analysis7,8,9.

尽管常用常规内镜检查方法，但它们依赖医学专业人员的经验和技能，再加上误诊的可能性，已不足以满足日益增长的诊断需求。因此，开发一种准确，快速且用户友好的检测胃肠道息肉的方法已成为当前医学研究的紧迫任务5,6。近年来，深度学习技术在计算机视觉领域取得了显着成功，为医学图像分析开辟了新的可能性7,8,9。

However, in the field of polyp detection, a series of challenges still exist, such as the multi-scale problem, structural complexity, and the diversity in the appearance of polyps.In addressing the multi-scale issue, various studies have proposed methods involving the fusion of multiple scale features10,11.

然而，在息肉检测领域，仍然存在一系列挑战，例如多尺度问题，结构复杂性和息肉外观的多样性。在解决多尺度问题时，各种研究提出了涉及多尺度特征融合的方法10,11。

However, these approaches have not entirely resolved the challenge of scale adaptability. Regarding structural intricacies, some research has introduced attention mechanisms to enhance the model’s focus on target structures12,13, yet capturing subtle differences remains a limitation. As for the diversity in the appearance of polyps, the current models still have limited generalization performance, making it difficult to effectively handle various complex polyp morphologies, which leads to suboptimal detection results in practic.

然而，这些方法并没有完全解决规模适应性的挑战。关于结构复杂性，一些研究引入了注意机制，以增强模型对目标结构的关注12,13，但捕捉细微差异仍然是一个局限性。至于息肉外观的多样性，目前的模型仍然具有有限的泛化性能，使得难以有效处理各种复杂的息肉形态，这导致实际中的检测结果不理想。

Designing the Channel-Guided Feature Integrator (CGFI) to enhance the model’s channel feature expression capabilities. This mechanism makes the network more flexible in adapting to different representations of target features;

设计信道引导特征积分器（CGFI）以增强模型的信道特征表达能力。这种机制使网络更灵活地适应目标特征的不同表示；

Introducing the Scale-Guided Aggregator (SGA) to aggregate information from different scales, enabling the network to comprehensively understand polyp target features. This enhances the model’s adaptability to representations of multi-scale targets, thereby significantly improving overall performance;.

引入规模引导聚合器（SGA）来聚合不同规模的信息，使网络能够全面了解息肉目标特征。这增强了模型对多尺度目标表示的适应性，从而显着提高了整体性能；。

Designing the Hierarchical Attention Integrator (HAI) to guide attention to crucial targets at different levels of the network. This hierarchical attention strategy aids in capturing subtle differences in target structures, enhancing the precision of polyp detection.

设计层次注意积分器（HAI），将注意力引导到网络不同层次的关键目标。这种层次注意策略有助于捕获目标结构的细微差异，提高息肉检测的精度。

Related workIn the early stages, the detection of gastrointestinal polyps primarily relied on advanced techniques such as endoscopic polyp detection and radiological detection. Through endoscopy, physicians can visually inspect the interior of the patient’s colon, facilitating the identification of polyps.

相关工作在早期阶段，胃肠道息肉的检测主要依赖于内镜息肉检测和放射学检测等先进技术。通过内窥镜检查，医生可以目视检查患者结肠的内部，有助于识别息肉。

Radiological detection encompasses traditional X-ray barium enema examinations and more sophisticated techniques like computerized tomography (CT) virtual colonoscopy, which utilize various imaging technologies for polyp detection.In recent years, with the rapid advancement of artificial intelligence and machine learning technologies, image processing and machine learning-assisted detection have emerged as research focal points to enhance the accuracy of traditional polyp detection14,15,16.

放射学检测包括传统的X射线钡灌肠检查和更复杂的技术，如计算机断层扫描（CT）虚拟结肠镜检查，该技术利用各种成像技术进行息肉检测。近年来，随着人工智能和机器学习技术的快速发展，图像处理和机器学习辅助检测已成为提高传统息肉检测准确性的研究重点14,15,16。

The application of deep learning technology in the field of polyp detection has made significant progress17,18,19, mainly due to its sophisticated algorithmic models. These models enable feature extraction and classification of medical images20,21, thereby improving the accuracy and efficiency of detection.

深度学习技术在息肉检测领域的应用取得了重大进展17,18,19，主要是由于其复杂的算法模型。这些模型能够对医学图像进行特征提取和分类20,21，从而提高检测的准确性和效率。

Currently, there are three main types of deep learning technologies in polyp detection: traditional Convolutional Neural Network (CNN) technology, two-stage object detection technology, and single-stage object detection technology.Traditional convolutional neural network methodConvolutional Neural Networks (CNNs) are among the popular structures in the field of deep learning, particularly well-suited for image recognition tasks.

目前，息肉检测中主要有三种深度学习技术：传统卷积神经网络（CNN）技术，两阶段目标检测技术和单阶段目标检测技术。传统的卷积神经网络方法卷积神经网络（CNN）是深度学习领域中流行的结构之一，特别适合于图像识别任务。

In the context of polyp detection, CNNs leverage their robust capabilities to automatically learn and extract crucial features from medical images, eliminating the need for tedious manual annotations or predefined feature extraction rules. This unique ad.

在息肉检测的背景下，CNN利用其强大的功能从医学图像中自动学习和提取关键特征，无需繁琐的手动注释或预定义的特征提取规则。这个独特的广告。

(1)

(1)

Next, within the channel attention mechanism, \(x_1\) undergoes an AvgPool (AP) operation, followed by another two-dimensional convolution (Conv2d) and a Sigmoid activation function. This process yields the channel attention weight map represented as \(x_2\), as expressed in Eq. (2):$$\begin{aligned} x_2=Sigmoid(Conv(AP(x_1))) \end{aligned}$$.

接下来，在通道注意机制中，\（x\U 1 \）经历AvgPool（AP）操作，然后是另一个二维卷积（Conv2d）和乙状结肠激活函数。这个过程产生了通道注意力权重图，表示为\（x\u 2 \），如等式（2）所示：$$\开始{对齐}x\u 2=乙状结肠（Conv（AP（x\u 1）））\结束{对齐}$$。

(2)

(2)

This weight map is utilized to dynamically adjust the importance of different channels in \(x_1\), allowing the network to focus more on channels critical for the target detection task.Finally, through element-wise multiplication, \(x_1\) is integrated with \(x_2\) to obtain the ultimate output y, As shown in Formula 3.

该权重图用于动态调整\（x\u 1 \）中不同通道的重要性，从而使网络更加关注对目标检测任务至关重要的通道。最后，通过元素乘法，将\（x\u 1 \）与\（x\u 2 \）积分以获得最终输出y，如公式3所示。

This integration process enables the network to flexibly focus on and utilize information from different channels of the input features, thereby enhancing the network’s perception and understanding of the target.$$\begin{aligned} y=x_1 * x_2 \end{aligned}$$.

这种集成过程使网络能够灵活地关注和利用来自输入特征不同渠道的信息，从而增强网络对目标的感知和理解$$\。

(3)

(3)

Throughout the entire process, Conv2d and BatchNorm operations contribute to linear transformations and normalization of features. The SiLU activation function introduces non-linear transformations, enhancing the expressive power of features. AvgPool and Sigmoid operations construct the channel attention mechanism, enabling the network to intelligently focus on different channels within the input feature map.By introducing this channel attention mechanism, the CGFI module effectively enhances the network’s focus on target features, providing the MCH-PAN framework with more powerful feature representation capabilities.

在整个过程中，Conv2d和BatchNorm操作有助于线性变换和特征归一化。SiLU激活函数引入了非线性变换，增强了功能的表达能力。AvgPool和Sigmoid操作构建了通道注意机制，使网络能够智能地关注输入特征图中的不同通道。通过引入这种通道注意机制，CGFI模块有效地增强了网络对目标特征的关注，为MCH-PAN框架提供了更强大的特征表示能力。

This is particularly evident in tasks involving multi-channel and multi-scale target detection, showcasing significant performance advantages.Scale-guided aggregatorThe Scale-Guided Aggregator (SGA) is one of the core components within the MCH-PAN framework. Its design aims to enhance the network’s adaptability to representations of targets at different scales.

这在涉及多通道和多尺度目标检测的任务中尤其明显，显示出显着的性能优势。规模引导聚合器规模引导聚合器（SGA）是MCH-PAN框架的核心组件之一。其设计旨在增强网络对不同规模目标表示的适应性。

This module effectively integrates features at different scales from the backbone network, including B2, B3, B4, and B5, to generate a more comprehensive and hierarchical feature representation. The pseudo-code for SGA is shown in Algorithm 2.Algorithm 2Scale-guided aggregatorFull size imageSGA first extracts corresponding feature representations from the features maps at four different scales, namely B2, B3, B4, and B5, from the backbone network, as shown in Eq.

该模块有效地集成了骨干网络中不同规模的功能，包括B2、B3、B4和B5，以生成更全面和分层的功能表示。SGA的伪代码如算法2所示。算法2Scale-guided aggregatorFull-size imageSGA首先从主干网络的四个不同尺度（即B2，B3，B4和B5）的特征映射中提取相应的特征表示，如等式2所示。

(4):$$\begin{aligned} Z=Concat(F_c(B_2)\sim F_c(B_5)) \end{aligned}$$.

（4） ：$$\开始{对齐}Z=连接（F\u c（B\u 2）\sim F\u c（B\u 5））\结束{对齐}$$。

(4)

(4)

Here, \(F_c\) refers to the CGFI module, and \(B_i\) refers to feature maps at different scales. These feature maps capture semantic information of the image at various scales, where B2 corresponds to coarser global information, while B5 corresponds to finer local details.Next, SGA introduces an intelligent scale-guidance mechanism, utilizing CGFI to dynamically aggregate these feature maps with weighted summation based on the scale characteristics of the target in the image, as depicted in Eq.

在这里，\（F\u c \）指的是CGFI模块，\（B\u i \）指的是不同尺度的特征图。这些特征映射以各种尺度捕获图像的语义信息，其中B2对应于较粗糙的全局信息，而B5对应于较精细的局部细节。接下来，SGA引入了一种智能比例引导机制，利用CGFI根据图像中目标的比例特征，通过加权求和来动态聚合这些特征图，如等式1所示。

(5):$$\begin{aligned} N=F_c(Z) \end{aligned}$$.

（5） ：$$\开始{对齐}N=F\u c（Z）\结束{对齐}$$。

(5)

(5)

This dynamic scale-guidance mechanism allows the network to flexibly attend to information at different scales, enabling the network to better adapt to the multi-scale representation of polyp targets in the image.Ultimately, the output of SGA is passed to the neck section of the model, specifically positions N3 and N4.

这种动态尺度引导机制允许网络灵活地处理不同尺度的信息，使网络能够更好地适应图像中息肉目标的多尺度表示。最终，SGA的输出被传递到模型的颈部，特别是位置N3和N4。

This provides the model with a more enriched and hierarchical feature representation, contributing to an enhanced sensitivity and detection performance for targets at different scales.Hierarchical attention integratorThe introduction of the Hierarchical Attention Integrator (HAI) is primarily aimed at establishing an effective attention mechanism between local features and high-dimensional features.

这为模型提供了更丰富和分层的特征表示，有助于提高不同尺度目标的灵敏度和检测性能。层次注意力整合层次注意力整合器（HAI）的引入主要旨在在局部特征和高维特征之间建立有效的注意机制。

The input to this module comes from the feature maps at positions N3, N4, and N5 in the model’s neck section. After processing by the CGFI, the output is then transmitted to higher layers, namely positions H4 and H5. The pseudo-code for SGA is shown in Algorithm 3.Algorithm 3Hierarchical attention integratorFull size imageIn the design of HAI, we first extract local features from the input, which encompass different hierarchical representations of the target, ranging from lower-level local information to higher-level semantic information.

该模块的输入来自模型颈部N3、N4和N5位置的特征映射。在CGFI处理后，输出随后被传输到更高层，即位置H4和H5。SGA的伪代码如算法3所示。算法3层次注意积分器全尺寸图像在HAI的设计中，我们首先从输入中提取局部特征，这些特征包含目标的不同层次表示，从低级局部信息到高级语义信息。

Subsequently, these local features are input to CGFI through a channel attention mechanism, dynamically weighting the information from different channels.The output of the channel attention mechanism is integrated with the original high-dimensional features, namely the feature maps at positions N3, N4, and N5 in the model’s neck section.

随后，这些局部特征通过通道注意机制输入到CGFI，动态加权来自不同通道的信息。通道注意机制的输出与原始的高维特征相结合，即模型颈部N3，N4和N5位置的特征映射。

This step aims to synthesize information from different levels, enabling the network to have a more comprehensive understanding of the target features. Ultimately, the integra.

这一步旨在综合不同层面的信息，使网络能够更全面地了解目标特征。最终，integra。

(6)

(6)

The introduction of the HAI module aims to enhance the model’s attention to local and high-dimensional features, strengthening the network’s expressive capabilities for targets through the channel attention mechanism. This hierarchical attention mechanism makes the model more adaptive, allowing it to better understand target information at different levels and scales.

HAI模块的引入旨在增强模型对局部和高维特征的关注，通过通道注意机制增强网络对目标的表达能力。这种分层注意机制使模型更具适应性，使其能够更好地理解不同级别和规模的目标信息。

Consequently, it provides a more powerful feature representation for target detection tasks.Loss functionThe loss function for MCH-PAN consists of the classification loss \(L_{cls}\) and the regression loss \(L_{reg}\). The regression loss \(L_{reg}\) is composed of the Complete Intersection over Union (CIoU) loss \(L_{CIoU}\) and the Distribution Focal Loss (DFL), as shown in Eq.

因此，它为目标检测任务提供了更强大的特征表示。损失函数MCH-PAN的损失函数由分类损失（L\uCls}）和回归损失（L\uReg}）组成。回归损失\（L\u{reg}\）由完全相交于联合（CIoU）损失\（L\u{CIoU}\）和分布焦点损失（DFL）组成，如等式所示。

(7):$$\begin{aligned} Loss=\lambda _\alpha L_{cls}+L_{reg}=\lambda _\alpha L_{cls}+\lambda _\beta L_{CIoU}+\lambda _\delta DFL \end{aligned}$$.

（7） ：$$\开始{对齐}损失=\λ\u \ alpha L\u{cls}+L\u{reg}=\λ\u \ alpha L\u{cls}+\λ\u \ beta L\u{CIoU}+\λ\u \ delta DFL \结束{对齐}$$。

(7)

(7)

Here, \(\lambda _\alpha\), \(\lambda _\beta\) and \(\lambda _\delta\) are different weight factors used to balance the importance of different losses.The classification loss is based on Variant Focal Loss (VFL), an improved cross-entropy loss function that employs different weighting schemes for positive and negative samples.

这里，\（\ lambda \ u \ alpha\），\（\ lambda \ u \ beta\）和\（\ lambda \ u \ delta\）是用于平衡不同损失重要性的不同权重因子。分类损失基于变异焦点损失（VFL），这是一种改进的交叉熵损失函数，对正样本和负样本采用不同的加权方案。

This asymmetrical weighting approach helps to better address the issue of sample imbalance. Additionally, VFL introduces an adaptive IoU weighting term to emphasize the importance of the anchor samples. When the sample is a positive one, this weighting term is adaptively adjusted based on the IoU value between the sample and the ground truth box, highlighting samples closer to the ground truth box, as shown in Eq.

这种不对称加权方法有助于更好地解决样本不平衡的问题。此外，VFL引入了一个自适应IoU加权项，以强调锚样本的重要性。当样本为正时，该加权项根据样本和地面真值框之间的IoU值进行自适应调整，突出显示更接近地面真值框的样本，如等式所示。

(8):$$\begin{aligned} VFL(p,q)={\left\{ \begin{array}{ll}-q(qlog(p)+(1-q)log(1-p))& q>0\\ alpha p^\gamma log(1-p)& q=0\end{array}\right. } \end{aligned}$$.

（8） ：$$\ begin{aligned}VFL（p，q）={\ left \{\ begin{array}{ll}-q（qlog（p）+（1-q）log（1-p））&q>0\\alpha p^ \ gamma log（1-p）&q=0 \结束{数组}\右。}\结束{对齐}$$。

(8)

(8)

The computation of the CIoU loss combines traditional IoU loss with additional geometric factors, including the distance between the center points of the predicted and ground truth boxes, as well as their aspect ratios. This comprehensive approach allows CIoU loss to more accurately reflect the differences between the predicted and ground truth boxes, guiding the model towards more precise localization, as illustrated in Eqs.

CIoU损失的计算将传统的IoU损失与其他几何因素相结合，包括预测和地面真值框的中心点之间的距离以及它们的纵横比。这种综合方法使CIoU损失能够更准确地反映预测和地面真值框之间的差异，从而引导模型实现更精确的定位，如方程式所示。

(9), (10):$$\begin{aligned} L_{CloU}=1-CIoU=1-(IoU-\frac{d_o^2}{d_c^2}-\frac{\nu ^2}{1-IoU+\nu }) \end{aligned}$$.

（9） ，（10）：$$开始{对齐}L\uu}=1-CIoU=1-（IoU-\ frac{d\u o ^ 2}{d\u c ^ 2}-\ frac{\nu ^ 2}{1-IoU+\nu}）\结束{对齐}$$。

(9)

(9)

$$\begin{aligned} \nu =\frac{4}{\pi ^2}(arctan\frac{w^{gt}}{h^{gt}}-arctan\frac{w}{h})^2 \end{aligned}$$

$$\ begin{aligned}\nu=\ frac{4}{\ pi ^ 2}（arctan \ frac{w ^{gt}}{h ^{gt}-arctan \ frac{w}{h}）^ 2 \ end{aligned}$$

(10)

(10)

In the equations, \(d_o\) represents the Euclidean distance between the centroids of the bounding box and the ground truth box, \(d_c\) is the diagonal distance between the bounding boxes, v is a parameter measuring aspect ratio consistency, \(w^{gt}\) and \(h^{gt}\) are the width and height of the ground truth box, w and h are the width and height of the predicted box, respectively.The primary idea behind DFL is to model the position of the box as a distribution.

在方程中，\（d\u o \）表示边界框质心与地面真值框之间的欧几里得距离，\（d\u c \）是边界框之间的对角线距离，v是衡量纵横比一致性的参数，\（w ^{gt}和\（h ^{gt}）是地面真值框的宽度和高度，w和h分别是预测框的宽度和高度。DFL背后的主要思想是将盒子的位置建模为分布。

By explicitly expanding the probabilities of \(y_i\) and \(y_{i+1}\) (\(y_i<y<y_(i+1)\)) , the network is forced to quickly focus on values near the label y, enhancing the accuracy of detection, especially when dealing with uncertain target positions. As shown in Eqs. (11–13):$$\begin{aligned} DFL\left( S_i,S_{i+1}\right) =-\left( \left( y_{i+1}-y\right) \log (S_i)+\left( y-y_i\right) \log (S_{i+1})\right) \end{aligned}$$.

通过显式扩展\（y\u i \）和\（y\u{i+1}\）（\（y\u i<y<y\u（i+1）\）的概率，网络被迫快速关注标签y附近的值，从而提高检测的准确性，特别是在处理不确定的目标位置时。如等式（11-13）所示：$$开始{对齐}DFL \左（S\U i，S\Ui+1}\右）=-\左（\左（y\Ui+1}-y \右）\ log（S\U i）+\左（y-y\U i \右）\ log（S\Ui+1}\右）\结束{对齐}$$。

(11)

(11)

$$\begin{aligned} S_i=\frac{y_{i+1}-y}{y_{i+1}-y_i} \end{aligned}$$

$$\开始{对齐}S\u i=\frac{y\u{i+1}-y}{y\u{i+1}-y\u i}\结束{对齐}$$

(12)

(12)

$$\begin{aligned} S_{i+1}=\frac{y-y_i}{y_{i+1}-y_i} \end{aligned}$$

$$\begin{aligned}S\uu{i+1}=\frac{y-y\u i}{y\uu{i+1}-y\u i}\end{aligned}$$

(13)

(13)

ExperimentDataset and experimental environmentThe dataset used in this study includes images of gastric and intestinal polyps, comprised of the Kvasir-SEG41 dataset and a gastric polyp dataset created by Zhang et al.33. In total, there are 1,758 endoscopic images containing polyp targets, which are divided into training, validation, and test sets with a ratio of 0.75:0.15:0.10.Although the Kvasir-SEG dataset provides bounding box information for polyp categories, this information cannot be directly used for training and validation of YOLO series models.

实验数据集和实验环境本研究中使用的数据集包括胃和肠息肉的图像，由Kvasir-SEG41数据集和Zhang等人创建的胃息肉数据集组成。。

To meet the requirements of YOLO series models, the bounding box information provided in the Kvasir-SEG dataset was extracted and normalized in conjunction with the image size. Some samples of the dataset and their labels are shown in Fig. 2, where the rectangular boxes indicate the detected polyps.Fig.

为了满足YOLO系列模型的要求，提取了Kvasir SEG数据集中提供的边界框信息，并结合图像大小进行了归一化。数据集的一些样本及其标签如图2所示，其中矩形框表示检测到的息肉。图。

2Dataset labeling.Full size imageThe dataset from Zhang et al.33 consists of images from 215 patients who underwent endoscopic examinations at Shao Yifu Hospital in Zhejiang Province, China, from January to June 2015. The dataset was augmented by rotating the original images by 180°. The Kvasir-SEG dataset41, on the other hand, was collected by the Vestre Viken Health Trust (VV) in Norway, which includes four hospitals providing medical services to a population of 470,000.

2数据集标签。全尺寸图像来自Zhang等[33]的数据集由2015年1月至6月在中国浙江省邵逸夫医院接受内镜检查的215名患者的图像组成。通过将原始图像旋转180°来增强数据集。另一方面，Kvasir SEG数据集41由挪威的Vestre Viken Health Trust（VV）收集，其中包括四家为47万人口提供医疗服务的医院。

The Bærum Hospital’s large gastroenterology department provided the training data. The combination of these two datasets enhances the model’s applicability across different populations and polyp targets.A detailed analysis of the processed dataset is provided in Figure x. Figure x (left) shows the distribution of bounding box center points, indicating that the targets are distributed in various directions, which is highly consistent with.

Bærum医院的大型胃肠科提供了培训数据。这两个数据集的组合增强了模型在不同人群和息肉目标中的适用性。图x（左）显示了边界框中心点的分布，表明目标分布在各个方向，这与高度一致。

(14)

(14)

Here, TP denotes True Positives, which are instances that are actually positive and correctly predicted as positive by the model. FP represents False Positives, indicating instances that are actually negative but incorrectly predicted as positive by the model.Recall refers to the proportion of instances that are actually positive among those predicted as positive by the model.

在这里，TP表示真正的积极因素，这些因素实际上是积极的，并且被模型正确预测为积极因素。FP表示假阳性，表示实际为阴性但模型错误预测为阳性的实例。回忆是指模型预测为阳性的实例中实际为阳性的实例的比例。

It is expressed by Eq.(15):$$\begin{aligned} Recall=\frac{TP}{TP+FN} \end{aligned}$$.

它由等式（15）表示：$$\ begin{aligned}Recall=\ frac{TP}{TP+FN}\ end{aligned}$$。

(15)

(15)

Here, FN represents False Negatives, which are instances that are actually positive but incorrectly predicted as negative by the model. A higher recall indicates that the model is able to correctly identify more positive instances among those that are actually positive, indicating a stronger detection capability of the model.The \(F1{\text{-}}Score\) is the harmonic mean of precision and recall, providing a balanced measure of a model’s performance.

在这里，FN表示假阴性，这是实际为正但模型错误预测为负的情况。。（F1{\ text{-}}得分）是精确度和召回率的调和平均值，为模型的性能提供了平衡的度量。

It is expressed by Eq. (16):$$\begin{aligned} F1{\text{-}}Score=\frac{2\times Precision\times Recall}{Precision+Recall} \end{aligned}$$.

它由等式（16）表示：$$\开始{对齐}F1{\文本{-}}得分=\分数{2倍精度\倍召回}{精度+召回}\结束{对齐}$$。

(16)

(16)

Mean Average Precision (mAP) is computed by calculating the Average Precision (AP) for each class at various recall levels and then taking the average across all classes. It is expressed by Eq. (17):$$\begin{aligned} mAP=\frac{1}{C}\sum _{i=1}^CAP_i \end{aligned}$$

平均平均精度（mAP）是通过计算每个类别在不同召回水平下的平均精度（AP），然后取所有类别的平均值来计算的。它由等式（17）表示：$$\ begin{aligned}mAP=\ frac{1}{C}\ sum{i=1}^ CAP\u i \ end{aligned}$$

(17)

(17)

Here, C represents the number of target classes, and \(AP_i\) denotes the Average Precision for the \(i-th\) class. The calculation of Average Precision (AP) is expressed by Eq. (18):$$\begin{aligned} AP=\int _0^1p(r)dr \end{aligned}$$

这里，C表示目标类的数量，\（AP\u i \）表示\（i-th \）类的平均精度。平均精度（AP）的计算由等式（18）表示：$$\ begin{aligned}AP=\ int{u 0 ^ 1p（r）dr \ end{aligned}$$

(18)

(18)

Here, p(r) represents precision at recall r. In practical computations, AP is often approximated using a discretized approach. This involves dividing the recall range from 0 to 1 into several intervals, calculating the maximum precision within each interval, and then averaging these maximum precisions to obtain AP.Comparative experimentTo assess the performance of different models under the same dataset and conditions, we conducted a series of comparative experiments.

这里，p（r）表示召回率r的精度。在实际计算中，AP通常使用离散化方法近似。这涉及将召回范围从0到1划分为几个区间，计算每个区间内的最大精度，然后对这些最大精度进行平均以获得AP。比较实验为了评估不同模型在相同数据集和条件下的性能，我们进行了一系列比较实验。

By comparing metrics such as Precision, Recall, \(F1{\text{-}}score\), and mAP (mean Average Precision) among different models, we can gain a more comprehensive understanding of the strengths and weaknesses of each model. This information serves as a valuable reference for subsequent research and practical applications.In the comparative experiments, this study evaluated several models, including earlier object detection models42,43,44,45, mainstream YOLO series models46,47,48, YOLO series models with task-specific improvements (YOLO-SPRD49, YOLO-OB50, Hyper-YOLO51, PATM-YOLO38, CAF-YOLO52) and MCH-PAN.

通过比较不同模型之间的精度，召回率，\（F1{\ text{-}}}得分\）和mAP（平均平均精度）等指标，我们可以更全面地了解每个模型的优缺点。这些信息为后续研究和实际应用提供了有价值的参考。在比较实验中，本研究评估了几种模型，包括早期的物体检测模型42,43,44,45，主流的YOLO系列模型46,47,48，具有特定任务改进的YOLO系列模型（YOLO-SPRD49，YOLO-OB50，Hyper-YOLO51，PATM-YOLO38，CAF-YOLO52）和MCH-PAN。

Each experiment was conducted three times, with the results presented as mean ± standard deviation. The experimental results are shown in Table 1.Table 1 Experimental results from different models.Full size tableFrom the results, earlier object detection models such as Faster R-CNN and RetinaNet achieved F1 scores of 0.635 ± 0.011 and 0.635 ± 0.009, and map@0.5:0.95 scores of 0.536 ± 0.003 and 0.533 ± 0.005, respectively.

每个实验进行三次，结果表示为平均值±标准偏差。实验结果如表1所示。表1来自不同模型的实验结果。全尺寸表根据结果，早期的物体检测模型（例如更快的R-CNN和RetinaNet）的F1得分分别为0.635±0.011和0.635±0.009，并且map@0.5：0.95分分别为0.536±0.003和0.533±0.005。

These scores are significantly lower compared to more advanced models. While these models perform reliably in general object detection tasks, they tend to fall short in complex medical imaging tasks such as polyp detection.Mainstream YOLO series models, such as YOLOv5s and .

与更先进的模型相比，这些分数要低得多。虽然这些模型在一般的物体检测任务中表现可靠，但在息肉检测等复杂的医学成像任务中往往表现不佳。主流YOLO系列车型，如YOLOv5s和。

Data availability

数据可用性

The data that support the fndings of this study are available from the corresponding author upon request.

支持本研究结果的数据可应要求从通讯作者处获得。

Code availability

代码可用性

The code for this study is available from the corresponding author upon request.

本研究的代码可应要求从通讯作者处获得。

ReferencesBelfiore, A. et al. Preventive anti-inflammatory diet to reduce gastrointestinal inflammation in familial adenomatous polyposis patients: A prospective pilot study. Cancer Prev. Res. 14, 963–972 (2021).Article

参考文献Belfiore，A。等人。预防性抗炎饮食减少家族性腺瘤性息肉病患者的胃肠道炎症：一项前瞻性初步研究。癌症预防。第14963-972号决议（2021年）。文章

CAS

中科院

Google Scholar

谷歌学者

Nguyen, L. H. et al. The sulfur microbial diet is associated with increased risk of early-onset colorectal cancer precursors. Gastroenterology 161, 1423–1432 (2021).Article

Nguyen，L.H.等人。硫微生物饮食与早发性结直肠癌前体的风险增加有关。胃肠病学1611423-1432（2021）。文章

CAS

中科院

PubMed

PubMed

Google Scholar

谷歌学者

Sninsky, J. A., Shore, B. M., Lupu, G. V. & Crockett, S. D. Risk factors for colorectal polyps and cancer. Gastrointest. Endosc. Clin. N. Am. 32, 195–213 (2022).Article

。胃肠学家。Endosc公司。临床。N、 第32页，195-213（2022）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Boland, C. R. et al. Diagnosis and management of cancer risk in the gastrointestinal hamartomatous polyposis syndromes: recommendations from the us multi-society task force on colorectal cancer. Gastroenterology 162, 2063–2085 (2022).Article

Boland，C.R.等人，《胃肠道错构瘤性息肉病综合征的癌症风险诊断和管理：美国多社会结直肠癌工作组的建议》。胃肠病学1622063-2085（2022）。文章

CAS

中科院

PubMed

PubMed

Google Scholar

谷歌学者

Eisenmann, M. et al. Biomedical image analysis competitions: The state of current participation practice. arXiv preprint arXiv:2212.08568 (2022).Brand, M. et al. Development and evaluation of a deep learning model to improve the usability of polyp detection systems during interventions.

Eisenmann，M.等人，《生物医学图像分析竞赛：当前参与实践的现状》。arXiv预印本arXiv:2212.08568（2022）。。

United Eur. Gastroenterol. J. 10, 477–484 (2022).Article .

欧洲胃肠病学联合会。J.10477-484（2022）。第条。

ADS

广告

Google Scholar

谷歌学者

Berzin, T. M. & Topol, E. J. Adding artificial intelligence to gastrointestinal endoscopy. Lancet 395, 485 (2020).Article

Berzin，T.M。＆Topol，E.J。将人工智能添加到胃肠道内窥镜检查中。柳叶刀395485（2020）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Ali, S. et al. Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge. Sci. Rep. 14, 2032 (2024).Article

。科学。代表142032（2024）。文章

ADS

广告

CAS

中科院

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Liaqat, A. et al. Gastric tract infections detection and classification from wireless capsule endoscopy using computer vision techniques: A review. Curr. Med. Imaging 16, 1229–1242 (2020).Article

。货币。医学成像161229-1242（2020）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Hasan, M. M., Islam, N. & Rahman, M. M. Gastrointestinal polyp detection through a fusion of contourlet transform and neural features. J. King Saud Univ. Comput. Inf. Sci. 34, 526–533 (2022).

Hasan，M.M.，Islam，N。＆Rahman，M.M。通过contourlet变换和神经特征的融合检测胃肠息肉。J、 。信息科学。34526-533（2022）。

Google Scholar

谷歌学者

Souaidi, M. & El Ansari, M. A new automated polyp detection network MP-FSSD in WCE and colonoscopy images based fusion single shot multibox detector and transfer learning. IEEE Access 10, 47124–47140 (2022).Article

Souaidi，M。＆El Ansari，M。一种新的自动息肉检测网络MP-FSSD，用于WCE和结肠镜检查图像，基于融合单次多盒检测器和转移学习。IEEE Access 1047124–47140（2022）。文章

Google Scholar

谷歌学者

Neto, A., Couto, D., Coimbra, M. T. & Cunha, A. Colonoscopic polyp detection with deep learning assist. In VISIGRAPP (4: VISAPP), 928–935 (2023).Zizhao, C. Hardnet-psa: a high precision polyp detection neural network with attention module. In 2021 IEEE International Conference on Emergency Science and Information Technology (ICESIT), 958–963 (IEEE, 2021).Misawa, M.

Cunha，A。结肠镜息肉检测与深度学习辅助。在VISIGRAPP（4：VISAPP），928-935（2023）中。Zizhao，C。Hardnet psa：一种具有注意力模块的高精度息肉检测神经网络。2021年IEEE国际应急科学与信息技术会议（ICESIT），958-963（IEEE，2021）。米萨瓦，M。

et al. Artificial intelligence-assisted polyp detection for colonoscopy: initial experience. Gastroenterology 154, 2027–2029 (2018).Article .

人工智能辅助结肠镜息肉检测：初步经验。胃肠病学1542027-2029（2018）。文章。

PubMed

PubMed

Google Scholar

谷歌学者

Fitting, D. et al. A video based benchmark data set (endotest) to evaluate computer-aided polyp detection systems. Scand. J. Gastroenterol. 57, 1397–1403 (2022).Article

Fitting，D.等人。一种基于视频的基准数据集（endotest），用于评估计算机辅助息肉检测系统。斯堪的纳维亚。J、 胃肠道。571397-1403（2022）。文章

CAS

中科院

PubMed

PubMed

Google Scholar

谷歌学者

Brand, M. et al. Frame-by-frame analysis of a commercially available artificial intelligence polyp detection system in full-length colonoscopies. Digestion 103, 378–385 (2022).Article

Brand，M.等人在全长结肠镜检查中对市售人工智能息肉检测系统进行逐帧分析。消化103378-385（2022）。文章

CAS

中科院

PubMed

PubMed

Google Scholar

谷歌学者

Wan, J., Chen, B. & Yu, Y. Polyp detection from colorectum images by using attentive yolov5. Diagnostics 11, 2264 (2021).Article

Wan，J.，Chen，B。＆Yu，Y。使用专注的yolov5从结直肠图像中检测息肉。诊断112264（2021）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Krenzer, A. et al. A real-time polyp-detection system with clinical application in colonoscopy using deep convolutional neural networks. J. Imaging 9, 26 (2023).Article

Krenzer，A。等人。使用深度卷积神经网络在结肠镜检查中具有临床应用的实时息肉检测系统。J、 成像9,26（2023）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Krenzer, A. et al. Fast machine learning annotation in the medical domain: a semi-automated video annotation tool for gastroenterologists. Biomed. Eng. Online 21, 33 (2022).Article

Krenzer，A。等人。医学领域的快速机器学习注释：胃肠病学家的半自动视频注释工具。生物医学。《工程在线》21，33（2022）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Krenzer, A., Hekalo, A. & Puppe, F. Endoscopic detection and segmentation of gastroenterological diseases with deep convolutional neural networks. In EndoCV@ ISBI, 58–63 (2020).Krenzer, A. et al. Automated classification of polyps using deep learning architectures and few-shot learning.

Krenzer，A.，Hekalo，A。＆Puppe，F。使用深度卷积神经网络对胃肠病学疾病进行内镜检测和分割。在EndoCV@ISBI，58-63（2020）中。。

BMC Med. Imaging 23, 59 (2023).Article .

BMC医学影像23,59（2023）。文章。

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Tajbakhsh, N., Gurudu, S. R. & Liang, J. Automated polyp detection in colonoscopy videos using shape and context information. IEEE Trans. Med. Imaging 35, 630–644 (2015).Article

Tajbakhsh，N.，Gurudu，S.R。＆Liang，J。使用形状和上下文信息在结肠镜检查视频中自动检测息肉。。医学影像35630-644（2015）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Urban, G. et al. Deep learning localizes and identifies polyps in real time with 96% accuracy in screening colonoscopy. Gastroenterology 155, 1069–1078 (2018).Article

。胃肠病学1551069-1078（2018）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Byrne, M. F. et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model. Gut 68, 94–100 (2019).Article

Byrne，M.F.等人。在使用深度学习模型分析标准结肠镜检查的未改变视频期间，腺瘤性和增生性小型结肠直肠息肉的实时分化。肠道68,94-100（2019）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Shin, Y., Qadir, H. A., Aabakken, L., Bergsland, J. & Balasingham, I. Automatic colon polyp detection using region based deep CNN and post learning approaches. IEEE Access 6, 40950–40962 (2018).Article

Shin，Y.，Qadir，H.A.，Aabakken，L.，Bergsland，J。＆Balasingham，I。使用基于区域的深度CNN和后学习方法自动检测结肠息肉。IEEE Access 640950–40962（2018）。文章

Google Scholar

谷歌学者

Mohammed, A., Yildirim, S., Farup, I., Pedersen, M. & Hovde, Ø. Y-net: A deep convolutional neural network for polyp detection. arXiv preprintarXiv:1806.01907 (2018).Ozawa, T. et al. Automated endoscopic detection and classification of colorectal polyps using convolutional neural networks.

穆罕默德，A.，Yildirim，S.，法鲁普，I.，佩德森，M。和霍夫德，Ø。Y-net：用于息肉检测的深度卷积神经网络。arXiv预印本XIV:1806.01907（2018）。Ozawa，T.等人。使用卷积神经网络对大肠息肉进行自动内镜检测和分类。

Ther. Adv. Gastroenterol. 13, 1756284820910659 (2020).Article .

他们。高级胃肠病。131756284820910659（2020）。文章。

Google Scholar

谷歌学者

Qadir, H. A. et al. Improving automatic polyp detection using CNN by exploiting temporal dependency in colonoscopy video. IEEE J. Biomed. Health Inform. 24, 180–193 (2019).Article

Qadir，H.A.等人，通过利用结肠镜检查视频中的时间依赖性，改进使用CNN的自动息肉检测。IEEE J.生物医学。健康信息。24180-193（2019）。文章

PubMed

PubMed

Google Scholar

谷歌学者

Qian, Z. et al. A new approach to polyp detection by pre-processing of images and enhanced faster R-CNN. IEEE Sens. J. 21, 11374–11381 (2020).Article

Qian，Z.等人。通过图像预处理和增强更快的R-CNN检测息肉的新方法。IEEE Sens.J.2111374–11381（2020）。文章

ADS

广告

Google Scholar

谷歌学者

Wang, R., Zhang, W., Nie, W. & Yu, Y. Gastric polyps detection by improved faster R-CNN. In Proceedings of the 2019 8th International Conference on Computing and Pattern Recognition, 128–133 (2019).Liu, Y., Ma, Z., Liu, X., Ma, S. & Ren, K. Privacy-preserving object detection for medical images with faster R-CNN.

Wang，R.，Zhang，W.，Nie，W。＆Yu，Y。通过改进的更快的R-CNN检测胃息肉。2019年第八届计算与模式识别国际会议论文集，128-133（2019）。Liu，Y.，Ma，Z.，Liu，X.，Ma，S。＆Ren，K。使用更快的R-CNN对医学图像进行隐私保护对象检测。

IEEE Trans. Inf. Forensics Secur. 17, 69–84 (2019).Article .

。Inf.法医学秘书。17,69-84（2019）。文章。

Google Scholar

谷歌学者

Sharma, V. et al. A multi-scale attention framework for automated polyp localization and keyframe extraction from colonoscopy videos. IEEE Trans. Autom. Sci. Eng. (2023).Zhang, X. et al. Real-time gastric polyp detection using convolutional neural networks. PLoS ONE 14, e0214133 (2019).Article .

Sharma，V。等人。用于自动息肉定位和结肠镜检查视频关键帧提取的多尺度注意框架。。自动。科学。工程（2023）。Zhang，X。等。使用卷积神经网络的实时胃息肉检测。PLoS ONE 14，e021133（2019）。文章。

CAS

中科院

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Chou, Y.-C. & Chen, C.-C. Improving deep learning-based polyp detection using feature extraction and data augmentation. Multimed. Tools Appl. 82, 16817–16837 (2023).Article

Chou，Y.-C.＆Chen，C.-C.使用特征提取和数据增强改进基于深度学习的息肉检测。多医学。工具应用。8216817-16837（2023）。文章

Google Scholar

谷歌学者

Shuang, L. et al. A real-time method for colorectal adenoma detection based on an improved yolov5s network. J. Hebei Univ. (Nat. Sci. Ed.) 42, 327 (2022).

双，L。等。基于改进的yolov5s网络的大肠腺瘤实时检测方法。河北大学学报（自然科学版）42327（2022）。

Google Scholar

谷歌学者

Liu, M., Jiang, J. & Wang, Z. Colonic polyp detection in endoscopic videos with single shot detection based deep convolutional neural network. IEEE Access 7, 75058–75066 (2019).Article

Liu，M.，Jiang，J。＆Wang，Z。基于单次检测的深度卷积神经网络在内镜视频中检测结肠息肉。IEEE Access 775058-75066（2019）。文章

PubMed

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Doniyorjon, M., Madinakhon, R., Shakhnoza, M. & Cho, Y.-I. An improved method of polyp detection using custom yolov4-tiny. Appl. Sci. 12, 10856 (2022).Article

Doniyorjon，M.，Madinakhon，R.，Shakhnoza，M.＆Cho，Y.-I.一种改进的使用定制yolov4-tiny的息肉检测方法。应用。科学。1210856（2022年）。文章

CAS

中科院

Google Scholar

谷歌学者

Wu, L. et al. Small gastric polyp detection based on the improved yolov5. Multimed. Tools Appl. 1–16 (2024).Lalinia, M. & Sahafi, A. Colorectal polyp detection in colonoscopy images using yolo-v8 network. Signal Image Video Process. 1–12 (2023).Liu, S., Qi, L., Qin, H., Shi, J. & Jia, J.

Wu，L。等人。基于改进的yolov5的小胃息肉检测。多医学。工具应用。1-16（2024年）。Lalinia，M。＆Sahafi，A。使用yolo-v8网络在结肠镜检查图像中检测结肠直肠息肉。信号图像视频处理。1-12（2023年）。刘S.，齐L.，秦H.，史J.＆Jia，J。

Path aggregation network for instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 8759–8768 (2018).Jha, D. et al. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, 451–462 (Springer, 2020).Ren, S., He, K., Girshick, R.

用于实例分割的路径聚合网络。在IEEE计算机视觉和模式识别会议论文集，8759-8768（2018）。Jha，D。等人。Kvasir seg：分段息肉数据集。。任，S.，何，K.，吉尔希克，R。

& Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Adv. Neural Inf. Process. Syst. 28 (2015).Lin, T.-Y., Goyal, P., Girshick, R., He, K. & Dollár, P. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, 2980–2988 (2017).Tan, M., Pang, R.

&Sun，J。更快的r-cnn：通过区域提议网络实现实时目标检测。高级神经信息处理。系统。28（2015）。Lin，T.-Y.，Goyal，P.，Girshick，R.，He，K。＆DolláR，P。密集物体检测的焦点损失。在IEEE国际计算机视觉会议论文集，2980-2988（2017）。Tan，M.，Pang，R。

& Le, Q. V. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10781–10790 (2020).Liu, W. et al. Ssd: Single shot multibox detector. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, 21–37 (Springer, 2016).Li, C.

&Le，Q.V。Efficientdet：可扩展且高效的对象检测。在IEEE/CVF计算机视觉和模式识别会议论文集，10781-10790（2020）。Liu，W。等人。Ssd：单次多盒探测器。《计算机视觉-ECCV 2016：第14届欧洲会议》，荷兰阿姆斯特丹，2016年10月11日至14日，《会议记录》，第一部分14、21至37（Springer，2016）。李，C。

et al. Yolov6: A single-stage object detection framework for industrial applications. arXiv preprintarXiv:2209.02976 (2022).Wang, C.-Y., Bochkovskiy, A. & Liao, H.-Y. M. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7464–7475 (2023).Wang, C.-Y., Yeh, I.-H.

等人。Yolov6：用于工业应用的单阶段对象检测框架。arXiv预印本XIV:2209.02976（2022）。Wang，C.-Y.，Bochkovskiy，A.＆Liao，H.-Y.M。约洛夫7号：可训练的免费礼品袋为实时物体探测器创造了最新技术。在IEEE/CVF计算机视觉和模式识别会议论文集，7464-7475（2023）。Wang，C-Y.，Yeh，I-H。

Google Scholar

谷歌学者

Yang, X. et al. Yolo-ob: An improved anchor-free real-time multiscale colon polyp detector in colonoscopy. arXiv preprintarXiv:2312.08628 (2023).Feng, Y. et al. Hyper-yolo: When visual object detection meets hypergraph computation. arXiv preprintarXiv:2408.04804 (2024).Chen, Z. & Lu, S.

Yang，X。等。Yolo ob：结肠镜检查中改进的无锚实时多尺度结肠息肉检测器。arXiv预印本XIV:2312.08628（2023）。。arXiv预印本XIV:2408.04804（2024）。Chen，Z。和Lu，S。

Caf-yolo: A robust framework for multi-scale lesion detection in biomedical imagery. arXiv preprintarXiv:2408.01897 (2024).Download referencesFundingThis research was supported, in part, by the National Natural Science Foundation of China under grant No.82302310, the Shenzhen Science and Technology Program under Grant JCYJ20220530172403008.Author informationAuthor notesThese authors contributed equally: Ling Wang, Jingjing Wan and Bolun Chen.Authors and AffiliationsFaculty of Computer and Software Engineering, Huaiyin Institute of Technology, Huaian, 223003, ChinaLing Wang, Xianchun Meng & Bolun ChenDepartment of Gastroenterology, The Second People’s Hospital of Huai’an, The Affiliated Huai’an Hospital of Xuzhou Medical University, Huaian, 223002, ChinaJingjing WanNanjing University of Aeronautics and Astronautics Shenzhen Research Institute, Shenzhen, 518038, ChinaWei ShaoAuthorsLing WangView author publicationsYou can also search for this author in.

Caf yolo：生物医学图像中多尺度病变检测的强大框架。arXiv预印本XIV:2408.01897（2024）。下载参考文献资助这项研究部分得到了国家自然科学基金82302310号资助，深圳科技计划JCYJ202530172403008号资助。作者信息作者注意到这些作者的贡献相同：王玲，万晶晶和陈伯伦。作者及所属单位淮安市淮阴工学院计算机与软件工程学院，223003，王玲，孟宪春，陈伯伦淮安市第二人民医院消化内科，徐州医科大学附属淮安医院，淮安，223002，南京航空航天大学深圳研究所，深圳，518038，中国邵伟作者凌旺维作者出版物您也可以在中搜索这位作者。

PubMed Google ScholarJingjing WanView author publicationsYou can also search for this author in

PubMed Google ScholarJingjing WanView作者出版物您也可以在

PubMed Google ScholarXianchun MengView author publicationsYou can also search for this author in

PubMed Google ScholarBolun ChenView author publicationsYou can also search for this author in

PubMed Google ScholarBolun ChenView作者出版物您也可以在

PubMed Google ScholarWei ShaoView author publicationsYou can also search for this author in

PubMed Google ScholarWei ShaoView作者出版物您也可以在

PubMed Google ScholarContributionsL.W. and B.C. design methodology, X.M. processing datasets, J.W. Conceived of the experiment, J.W. and L.W. conducted the experiment, L.W. and W.S. Results were analyzed, L.W. and J.W. wrote the first draft, B.C. and W.S. reviewed and edited.

PubMed谷歌学术贡献l。W、 和B.C.设计方法论，X.M.处理数据集，J.W.构思了实验，J.W.和L.W.进行了实验，分析了L.W.和W.S.的结果，L.W.和J.W.撰写了初稿，B.C.和W.S.进行了审查和编辑。

All authors reviewed the manuscript.Corresponding authorsCorrespondence to.

。通讯作者通讯。

Ling Wang, Jingjing Wan or Wei Shao.Ethics declarations

王玲，万晶晶或魏绍。道德宣言

Competing interests

相互竞争的利益

The authors declare no competing interests.

作者声明没有利益冲突。

Additional informationPublisher’s noteSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissions

Additional informationPublisher的noteSpringer Nature在已发布地图和机构隶属关系中的管辖权主张方面保持中立。权限和权限

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 articleWang, L., Wan, J., Meng, X. et al. MCH-PAN: gastrointestinal polyp detection model integrating multi-scale feature information.

转载和许可本文引用本文Wang，L.，Wan，J.，Meng，X。et al。MCH-PAN：整合多尺度特征信息的胃肠息肉检测模型。

Sci Rep 14, 23382 (2024). https://doi.org/10.1038/s41598-024-74609-9Download citationReceived: 01 April 2024Accepted: 27 September 2024Published: 08 October 2024DOI: https://doi.org/10.1038/s41598-024-74609-9Share 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.

科学报告1423382（2024）。https://doi.org/10.1038/s41598-024-74609-9Download引文收到日期：2024年4月1日接受日期：2024年9月27日发布日期：2024年10月8日OI：https://doi.org/10.1038/s41598-024-74609-9Share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

CommentsBy submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

评论通过提交评论，您同意遵守我们的条款和社区指南。如果您发现有虐待行为或不符合我们的条款或准则，请将其标记为不合适。