3cDe-Net： A Cervical Cancer Cell Detection Network Based on an Improved Backbone Network and Multiscale Feature Fusion

3cDe-Net： A Cervical Cancer Cell Detection Network Based on an Improved Backbone Network and Multiscale Feature Fusion | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article 3cDe-Net： A Cervical Cancer Cell Detection Network Based on an Improved Backbone Network and Multiscale Feature Fusion Wei Wang, Yun Tian, Yang Xu, Xiao-xuan Zhang, Yan-song Li, Shi-feng Zhao This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-1485286/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Background Cervical cancer cell detection is an essential means of cervical cancer screening. However, in thin-prep cytology test-based images, the detection accuracy of traditional computer-aided detection algorithms is typically low due to the overlap of cells with blurred cytoplasmic boundaries. Clinical applications face more difficulties. Method we propose a cervical cancer cell detection network (3cDe-Net) based on an improved backbone network and multiscale feature fusion, and the network consists of the backbone network and the detection head. In the backbone network, a dilated convolution and a group convolution are introduced to improve the resolution and model expression ability. In the detection head, multiscale features are obtained based on a feature pyramid fusion network to ensure accurate capture of small cells; and then, based on Faster R-CNN, adaptive anchors of cervical cancer cells are generated via unsupervised clustering; furthermore, a new balanced L1-based loss function is defined, which reduces the unbalanced sample contribution loss. Result For two different datasets(the Data-T dataset and Herlev dataset), the baselines including ResNet-50, ResNet-101, Inception-v3, ResNet-152 and the feature concatenation network were used, and the final quantitative results showed the effectiveness of the proposed backbone network DC-ResNet. Furthermore, experiments with both datasets show that our detection network 3cDe-Net, based on the optimal anchor, the defined new loss function, and the DC-ResNet, outperforms existing methods and achieves a mAP of 50.4%. By horizontal comparison of cells on the image, the category information and location information can be obtained concurrently. Conclusion The proposed 3cDe-Net is an end-to-end network that can detect cancer cells on multicell pictures and can determine their locations. The model processes and analyzes samples directly at the picture level rather than the cellular level, which is more efficient and meets clinical needs more effectively. Cervical cancer detection Feature fusion Backbone network Adaptive anchors， Loss function Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction Cervical cancer is the fourth most common gynecological malignancy globally. In 2018, there were approximately 570,000 new cases and 310,000 deaths worldwide [ 1 ]. Traditional cervical cancer cell screening typically requires a pathologist to observe thousands of cells under a microscope and provide a report based on diagnostic criteria[ 2 ]. This approach is time-consuming, labor-intensive, relies heavily on the doctor's experience and has a strong subjectivity[ 3 ]. Computer-aided cervical cancer cell detection is likely to become common for clinical diagnosis to solve the above problems [ 4 , 5 ]. According to whether the segmentation step is included in the analysis pipeline, the identification of cervical cancer cells can be divided into segmentation-based recognition and objection detection-based recognition. Segmentation-based recognition methods typically segment cells or cell components and then extract cell characteristics for cell classification [ 6 – 10 ]. Detection accuracy typically depends on the results of cell segmentation, which makes it difficult to accurately identify overlapping cells with blurred cytoplasmic boundaries. Clinical applications face more difficulties. Object-detection-based cervical cancer cell recognition has been a trend in recent years[ 11 – 13 ] and use an object detection framework based on a convolutional neural network to obtain the classification and location of cancer cells. According to whether regional candidate boxes are generated, object detection methods can be divided into two categories: two-stage methods and one-stage methods[ 14 ]. Xiang et al.[ 12 ] proposed an automatic assisted cervical cell screening system based on YOLO-v3-net and designed a classifier to further distinguish the categories of hard samples. Zhuang et al.[ 13 ] designed a special backbone network for cervical cancer cells and applied it to the single shot multibox detector (SSD) framework. These deep learning-based cervical cancer cell detection algorithms are one-stage [ 12 , 13 ], and their detection accuracy is not high. Xu et al.[ 11 ] proposed a two-stage detection method and transplanted the Faster R-CNN[ 15 ] framework of natural images into cervical cancer cells. However, the difference between cervical cancer cells and natural images is not considered; as a result, the detection performance is weak. In this study, we propose a two-stage cervical cancer detection algorithm based on an improved backbone network and multiscale feature fusion. In the first stage, cervical cell features with different sizes are extracted via an improved backbone network, and feature extraction can be verified through classification experiments. In the second stage, the location information of cervical cancer cells is obtained via the detection network. The features with different scales are fused through the feature pyramid, and then, the adaptive anchors are located by K-means clustering. Also, a loss function that alleviates the imbalance of sample contribution is defined, which improves detection accuracy. The proposed approach processes and analyzes samples directly at the picture level rather than the cellular level, which is more efficient and meets clinical needs more effectively. By horizontal comparison of cells on the image, the category information and location information can be obtained concurrently. The primary contributions of this study are as follows: To address the problem of different cell sizes, we present a two-stage cervical cancer detection method based on multiscale feature fusion. We introduce dilated convolution and group convolution in the backbone into a deep residual network. The proposed 3cDe-Net avoids upsampling operations and reduces the information loss of small cells on the feature map; Due to the characteristics of cervical cancer cells, K-means clustering is used to determine the anchor frame size and ratio that are more suitable for cervical cancer cells and can providing better prior knowledge. We also define a loss function to address the imbalance between negative and positive samples for cervical cancer cell detection. 2. Related Works 2.1 Backbone network Currently, deep learning-based detection algorithms must typically identify the features of an input picture through the feature extraction network. The feature extraction network of the classification task is also known as the backbone network. VGGNet[ 16 ] is the backbone network of Faster R-CNN[ 15 ], and the structure is simple. ResNet uses a deeper network structure to extract more complex features[ 17 ]. Both networks are still relatively common backbone networks. In addition, DenseNet densely connects each layer in the network with other layers, as well as DetNet specifically designed for object detection[ 18 , 19 ]. Existing backbone networks are primarily used to recognize natural images. However, in cervical cell images, the canceration of cells is a gradually changing process, which makes it difficult to distinguish normal cells from cancerous cells using traditional backbone networks. Also, the scale of cervical cells varies, and small cells are difficult to identify on deeper feature maps, which makes it more difficult to detect these small cells. Thus, we use group convolution to enhance the expression ability of the extracted features. Concurrently, to better distinguish normal cells from abnormal cells, a dilated convolution is used to improve the resolution of the generated map, as well as the classification accuracy of small cells. Furthermore, this convolution also reduces the number of calculations caused by upsampling during feature map fusion. 2.2 Object detection based on deep learning Compared with classification, detection has an additional location task. Therefore, based on the backbone network, a network called the detection head should be added to locate the object region proposal. Thus, the backbone network and the detection head together construct the detection network. According to whether the detection head contains the region proposal network, the detection networks can be divided into one- and two-stage methods. The one-stage detection methods do not generate the region proposals, and the location and category prediction are completed in one stage. However, the accuracy of these methods could be improved. The two-stage methods first perform pre-detection by generating regional proposals and then fine-tuning the location and classification, yielding high accuracies. Faster R-CNN [ 15 ] is a classic two-stage detection network and is effective for many natural image datasets[ 20 ], but the network is not suitable for cervical cells due to the anchors generated based on natural images. In addition, the small cells retain less information on the feature map, which affects the detection accuracy of these cells. Finally, the loss function of the network does not consider the unbalanced sample size. To solve these problems, we use feature pyramids to integrate deep and shallow feature maps, which integrates deep semantic features and shallow location information more effectively. Therefore, small cells can be detected, and adaptive anchor boxes can be generated. Furthermore, a new loss function is defined, and the detection accuracy of cervical cancer cells is improved. 3. Proposed Methodology The overall framework of the proposed network is shown in Fig. 1 . The backbone network uses a dilated convolution ResNet (DC-ResNet). The algorithm first performs pre-detection of the cancer cells through the region proposal network (RPN) and then obtains the results through the classification and regression network. Multiple feature maps with different scales are generated from the backbone network. Feature fusion is performed via the pyramid network to obtain the final predicted feature map, and then, the feature map is fed to the region proposal network. The feature pyramid network (FPN) achieves feature fusion by upsampling to make the depth and shallow feature maps with the same dimension. The feature pyramid can transfer deep semantic features to shallow layers to supplement semantic information. As a result, high-resolution and strong semantic features are obtained, which are fit to detect small objects. The RPN first adaptively generates anchor boxes on the generated prediction feature maps and then selects and adjusts the anchor boxes to obtain better region proposals. Next, the proposals and feature maps are fed into the classification and regression network. Finally, cervical cancer cells are predicted and located. 3.1 Improved Backbone Network: DC-ResNet DC-ResNet derived from ResNet is an improved backbone network. In DC-ResNet, a dilated convolution and a group convolution are introduced, and the details of the network structure is shown in Fig. 2 . The input images are fed to the network successively through the convolutional layer (Conv), batch normalization layer (BN), ReLU activation function and the pooling layer (MaxPooling), and then, the feature maps are obtained via multiple convolution groups. The first three groups (blue) use residual grouping convolution, and the last two groups (yellow) use residual dilated convolution. To improve the feature expression ability of the network, the feature map is fed into the fully connected layer to obtain the score of the predicted category. Each fully connected layer is followed by a dropout layer to prevent overfitting. Grouped convolution was originally a training method[ 21 ] that was designed to solve hardware resource limitations. To obtain more distinguishable features of cervical cancer cells, we group convolutional divisions on the channels, and then, the results of each group are concatenated . The hyperparameter problem is solved by grouping convolution. Thus, model accuracy is improved without increasing the number of parameters. The convolution operation is performed by multiple GPUs, and the calculation results are connected. The calculation process is shown in Fig. 3 , where \(c\) is the dimension before convolution, and \(d\) is the dimension after convolution. In this study, residual group convolution is introduced, which is the introduction of group convolution based on residual networks, and the details are shown in Fig. 4 . The left panel is the overall structure diagram, and the right panel is the detailed diagram of the grouped convolution. Residual hole convolution uses hole convolution on the residual network. The calculation formula of the output feature size of the hole convolution is as follows: $$\begin{array}{c}y=⌊\frac{x-k-\left(k-1\right)\times \left(d-1\right)+2p}{s}⌋+1\#（1）\end{array}$$ where \(y\) is the size of the output feature map, \(x\) is the size of the input image, \(k\) is the size of the convolution kernel, \(d\) is dilation, \(p\) is padding, and \(s\) is stride. The feature maps before and after convolution have the same parameters. The residual dilated convolution structure is shown in Fig. 5 . The overall structure is similar to the residual group convolution, but the dilated convolution is used instead of the group convolution. The size of the feature map after the dilated convolution is unchanged. The residual dilated convolution has two structures A and B. The difference is whether the residual branch contains added 1×1 convolution. Downsampling/upsampling operations are not required for feature fusion. Thus, the size of the image feature map is larger than that of the original image. Thus, small cells are more informative at each feature point, and more edge information is retained. 3.2 Generation of anchor box When RPN generates region proposals, anchor boxes with different sizes on different feature layers of FPNs can be generated. The size and ratio of the anchor boxes are typically set based on domain prior knowledge or datasets. However, in this study, the changes in the size and the ratio of cancer cells were large, which is a challenge for anchor box generation. Inspired by YOLO v2[ 22 ], the best \(k\) boxes can be obtained by K-means clustering. The locations of all cells are not certain. We measure whether the size and ratio of anchors by looking at all the target boxes as being at the origin position. When clustering is performed, the distances are calculated between predicting boxes and centers by Eq. (2): $$\begin{array}{c}b\left(box,center\right)=1-IoU\left(box,center\right)\#\left(2\right)\end{array}$$ where IoU represents the fitting degree of the two boxes and is the ratio of the area of the intersection and the union of the prediction box and the real box, as shown in Fig. 6 . The deep feature maps are small, and the receptive field is large, which is good for large cells. The shallow feature maps are the opposite, which is more suitable for detecting small cells. Therefore, anchors with different sizes and proportions can be generated on each feature map. 3.3 Definition of the loss function The detection network is a multitask learning model that must predict the classification and location of cervical cancer cells. Therefore, the loss function should be considered to include classification and regression, and the definition is as follows: $$\begin{array}{c}L\left(\left\{{P}_{i}\right\},\left\{{bbox}_{i}\right\}\right)=\frac{1}{{N}_{cls}}\sum _{i}{L}_{cls}\left({p}_{i},{p}_{i}^{*}\right)+\lambda \frac{1}{{N}_{reg}}\sum _{i}{P}_{i}^{*}{L}_{reg}\left({bbox}_{i},{bbox}_{i}^{*}\right)\#\left(3\right)\end{array}$$ where \(\sum _{i}{L}_{cls}\left({p}_{i},{p}_{i}^{*}\right)\) is the classification loss, \({p}_{i}\) is the real category, and \({p}_{i}^{*}\) is the predicting category. The classification function is calculated using cross-entropy loss, and \(\lambda\) is the weight that balances the two task losses. \(\sum _{i}{P}_{i}^{*}{L}_{reg}\left({bbox}_{i},{bbox}_{i}^{*}\right)\) represents regression loss and only calculates positive samples, not including negative samples. The \({smooth}_{L1}\) function is used to calculate the regression loss. The definitions are as follows: $$\begin{array}{c}{L}_{reg}\left({bbox}_{i},{bbox}_{i}^{*}\right)=\sum _{i\in x,y,w,h}{smooth}_{L1}\left({bbox}_{i}-{bbox}_{i}^{*}\right)\#\left(4\right)\end{array}$$ $$\begin{array}{c}{smooth}_{L1}\left(x\right)=\left\{\begin{array}{cc}0.5{x}^{2}& if\left|x\right|<1\\ \left|x\right|-0.5& otherwise\end{array}\right.\#\left(5\right)\end{array}$$ Setting the weights \(\lambda\) to balance classification and the positioning loss remains challenging. The parameter \(\lambda\) is usually set manually. However, when calculating the regression task of the position, the model will consider the samples with a regression loss greater than 1 more when λ increases because the loss of the regression task is unconstrained. Therefore, when designing the loss function, more consideration should be given to those samples with a loss of less than 1. Inspired by Pang et al.[ 23 ], the original \({smooth}_{L1}\) is replaced with \({balanced}_{L1}\) defined as follows: $$\begin{array}{c}{balanced}_{L1}\left(x\right)=\left\{\begin{array}{cc}\frac{\alpha }{b}\left(b\left|x\right|+1\right)\text{ln}\left(b\left|x\right|+1\right)& if\left|x\right|<1\\ \gamma \left|x\right|+C& otherwise\end{array}\right.\#\left(6\right)\end{array}$$ where α is used to control the gradient change of samples with a loss of less than 1 and γ is used to adjust the upper limit of the error. By adjusting the above two parameters, it is possible to balance the gradient contribution of each sample. 4. Experiments And Results 4.1 Datasets and evaluation metrics The experimental data evaluated in this study are from the Tian-chi competition datasets (Data-T) [1 ] and Herlev [2] . Fig. 7 shows several images from the dataset. In the Tian-chi competition dataset, each cervical cell smear image contains multiple cervical cells, which can be used for classification and detection. In the Herlev image dataset, each image contains only one cervical cell, which can be used for only classification. Data-T: This dataset comes from the preliminary data of the Cervical Cancer Risk Diagnosis Intelligent Challenge and contains 800 images of the Thinprep cytologic test (TCT) labeled by a professional pathologist, including 500 positive images and 300 negative images. Positive pictures were used to label the location of abnormal squamous epithelial cells. Due to the large size of the original pathological image (each image is approximately 40835 ×42371 pixels), we divided the original pathological image into several images with 800×800 pixels to facilitate processing. Thus, there were 6,627 abnormal squamous epithelial cells. A negative sample refers to an image that does not contain cervical cancer cells. Dataset samples only labeled the location of abnormal squamous epithelial cells. In this study, 6627 normal squamous epithelial cells were screened from 300 negative pictures using semisupervised learning methods, and together with 6627 positive samples, a total of 13254 sample images were described by a classification dataset of cervical cancer cells. Images were divided into training sets, validation sets and test sets according to a ratio of 8:1:1, and the ratio of positive to negative samples was 1:1. To train the model with good generalizability, we enhanced the sample images using operations including rotation transformation and flip transformation. Herlev: This dataset contains images of cervical cancer cells that were collected by Herlev University Hospital in Denmark. There are 917 single-cell images with 200×100 pixels, including 242 normal cells and 675 abnormal cells. This dataset has become the primary study dataset for the classification of cervical cancer cells. Due to the small amount of data and the imbalance between the positive and negative samples, we first augmented the sample images by center rotation and translation operations. Normal cell images were rotated 20 times, and abnormal cell images were rotated 10 times. As a result, the number of positive and negative samples is approximately equal. Finally, the training set, verification set, and test set were divided by a ratio of 8:1:1. In the classification experiments, metrics including sensitivity, specificity, h-mean, f1 measure and accuracy were used to evaluate the performance of the proposed feature extraction network. Sensitivity represents the proportion of correct images in all predicted cancer cell images, and specificity represents the proportion of correct images in all predicted normal cells. The detection experiment was evaluated by the mean average precision (mAP). 4.2 Network parameters and implementation details The size of the input images of the backbone network is 224×224×3, and the details of the network structure and parameters of DC-ResNet are listed in Table 1. Experiments were performed on a workstation with the Ubuntu16.04 operating system and an NVIDIA GeForce 2080Ti 12 GB GPU. While training the backbone network, the SGD optimization algorithm was used to optimize the model parameters. The batch size was set to 32. The learning rate of each layer was initially set to 0.01. After 50 epochs of training, the learning rate was reduced to 1/10 every 10 epochs. The momentum was set to 0.9, and training ended after 1000 epochs. While training the detection network, the SGD optimization algorithm was used to optimize the model parameters. The batch size was set to 6, the learning rate of each layer was initially set to 0.00125, and the learning rate was reduced to 1/10 after 16 epochs and 22 epochs. The momentum was set to 0.9, and the weight was decayed by 0.001. Table 1 Structure and parameters of DC-ResNet Improved backbone: DC-ResNet \(7\times 7\) , 64, stride 2 \(3\times 3\) , max pool, stride 2 Residual group convolution \(\left[\begin{array}{cc}1\times 1& 128\\ 3\times 3& 128\\ 1\times 1& 256\end{array}\right]\times 3\) Group 32 \(\left[\begin{array}{cc}1\times 1& 256\\ 3\times 3& 256\\ 1\times 1& 512\end{array}\right]\times 4\) Group 32 \(\left[\begin{array}{cc}1\times 1& 512\\ 3\times 3& 512\\ 1\times 1& 1024\end{array}\right]\times 6\) Group 32 Residual dilated convolution \(B：\left[\begin{array}{cc}1\times 1& 1024\\ 3\times 3& 256\\ 1\times 1& 1024\end{array}\right]\times 1\) \(A:\left[\begin{array}{cc}1\times 1& 1024\\ 3\times 3& 256\\ 1\times 1& 1024\end{array}\right]\times 2\) dilation 2 stride 2 \(B：\left[\begin{array}{cc}1\times 1& 1024\\ 3\times 3& 256\\ 1\times 1& 1024\end{array}\right]\times 1\) \(A:\left[\begin{array}{cc}1\times 1& 1024\\ 3\times 3& 256\\ 1\times 1& 1024\end{array}\right]\times 2\) dilation 2 stride 2 fc-1024 fc-256 fc-2 4.3 Results and analysis With the Data-H dataset, we used ResNet-50 and ResNet-101 as baselines for comparison with the proposed DC-ResNet. Table 2 lists the quantitative comparison results and shows that the proposed backbone network (DC-ResNet) performs better than baselines in all other metrics except for sensitivity. However, specificity is more important than sensitivity in the detection of cervical cancer cells due to the majority of normal cervical cell samples. The proposed DC-ResNet has 59 convolutional layers, and ResNet-50 has 50 convolutional layers. To verify the validity of the network structure of the proposed DC-ResNet, we compared DC-ResNet with ResNet-101 with 101 convolutional layers. Table 2 shows that DC-ResNet outperforms the other models. Also, all evaluation metrics, except for sensitivity, of the ResNet-101 network are lower than those of ResNet-50, which may be due to a limited number of datasets. Although no fittings of these complex models occurred, they do not necessarily produce better results. Table 2 Quantitative comparison on the Data-H dataset Method H-means Sensitivity Specificity F1 Accuracy ResNet-50 96.82% 96.68% 96.98% 96.82% 96.83% ResNet-101 96.75% 97.12% 96.37% 96.76% 96.75% DC-ResNet 97.11% 95.92% 98.34% 97.09% 97.13% With the Herlev dataset, we used Inception-v3[25], ResNet-152[17] and a feature concatenation network[24] as baselines for comparison with the proposed DC-ResNet. The details of the quantitative comparison are listed in Table 3, which shows that the proposed DC-ResNet achieves the highest classification accuracy by nearly 4%. Due to the small amount of data in the Herlev dataset, the 5-fold cross-validation method was used to verify the proposed network. Results indicate that the proposed DC-ResNet is superior to the baselines in terms of accuracy and exhibits better stability. The partial recognition results of DC-ResNet on the two datasets are shown in Fig. 8. Table 3 Quantitative comparison on the Data-H dataset Method Accuracy Inception-v3 [25] 89.66±1.89% Resnet-152 [17] 90.87±1.48% Feature concatenation [24] 92.63±1.68% DC-ResNet 96.7%±1.1% To verify the effectiveness of DC-ResNet in the detection of cervical cancer cells, we performed a comparative experiment with ResNet-50 and ResNet-101 on mAP indicators, and the results are listed in Table 4. The mAP of ResNet-101 is not higher than that of ResNet-50, which shows that simply increasing the complexity of the model may lead to model overfitting with small datasets. Concurrently, the mAPs of DC-ResNet-50 are at least 1% higher than those of ResNet-50, which shows that the improved performance of the proposed model is not due to the increase in the number of network layers but rather the effect of changes in the network structure. Table 4 mAP results of different backbone networks Backbone [email protected] （ % ） [email protected] （ % ） ResNet-50 45.4 26.2 ResNet-101 45.5 25.9 DC-ResNet 46.7 26.5 We also analyzed the influence of the number of feature maps of DC-ResNet. F[1,2,3,4], F[1,2,3,5], F[1,2,4,5] and F[1,2,3,4,5] represent the feature fusion layers of the PFN of the corresponding feature layers in Fig. 1. The detection effect of feature fusion with different combinations is shown in Table 5. The worst detection effect is 46% when the input feature map is numbered [1,2,3,4]. Although the number of feature maps is reduced, the mAP is still 5% higher than ResNets. This result indicates that the superior performance of the proposed model is due to the increase in the number of feature layers and the structure of the proposed DC-ResNet itself. Table 5 Results of FPN fusion of different layer feature maps The number of fusion feature map [email protected] (%) 1，2，3，4 46.0 1，2，3，5 46.5 1，2，4，5 46.1 1 ， 2 ， 3 ， 4 ， 5 46.7 Table 6 shows the mAPs of the detection network with different improvement measures. The performance of the proposed 3cDe-Net is shown to be the best, with a maximum mAP of 50.4%. The reasons for the performance improvement of the proposed model are the anchor obtained by K-means clustering, the improved loss function, and DC-ResNet replacing the backbone network. The mAP metric increased by 3% due to the improvements in the detection network and the backbone network. The average IoU increased by 5% due to the anchor size generated by K-means clustering, and the mAP increased by 0.8% by introducing the generated anchor ratio into the detection network. In the improved , the mAP increased by 1.2% with and . By replacing the backbone network with DC-ResNet from the original ResNet-50, the mAP increased by 1.1%. Several detection results of our 3cDe-Net are shown in Fig. 9. Table 6 Detection results of our 3cDe-Net Improved anchor Improved Loss DC-ResNet [email protected] （%） 47.3 √ 48.1 √ √ 49.3 √ √ √ 50.4 1 https://tianchi.aliyun.com/competition/entrance/231757/introduction 2 http://mde-lab.aegean.gr/index.php/downloads 5. Conclusion Based on the characteristics of cervical cancer cells, we proposed a multiscale feature fusion-based cervical cancer cell detection network based on an improved the backbone network, which contains a backbone network and a detection head. The backbone network was improved by group convolution and dilated convolution to optimize feature extraction, which increases the network receptive field while effectively downsampling. As a result, small cell features are more convenient to express. The detection head based on Faster R-CNN was improved, and different anchor boxes could be automatically set for different target sizes. Also, targets with different sizes were able to be predicted on different depths of feature layers. To minimize regression loss, a new balanced L1-based loss function was developed to reduce unbalanced sample contribution losses. Experiments with the Data-T and Herlev datasets show that the proposed model outperforms existing methods and achieves a mAP of 50.4%. The next step in this field of research is to identify the type and stage of cancer cells based on the identification of cervical cancer cells. Declarations Ethics approval and consent to participate We confirm that all methods were carried out in accordance with relevant guidelines and regulations. Consent for publication Not applicable Availability of data and materials The datasets analysed during the current study are available in the Tian-chi Initiative and Herlev. https://tianchi.aliyun.com/competition/entrance/231757/introduction http://mde-lab.aegean.gr/index.php/downloads Conflicts of Interest The authors declare that there is no conflict of interest regarding the publication of this paper. Funding This work has been partially supported by the National Natural Science Foundation of China (No. 62172047 and No. 61802020) and the Major Program of National Natural Science Foundation of China (No. 72091511). Acknowledgements The authors would like to express appreciation to the anonymous reviewers and editors for their helpful comments that improved the paper, and also would like to thank American Journal Experts for providing linguistic assistance during the preparation of this paper. Authors' information 1 Key laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Department of Gynecologic Oncology, Peking University Cancer Hospital & Institute, Beijing,100142, China; 2 School of Artificial Intelligence, Beijing Normal University, Beijing, 100875, China. Correspondence should be addressed to Yun Tian; [email protected] Authors’ contributions WW, XxZ, and XY conceived the study and wrote the manuscript. XY, XxZ, and YsL performed training of convolutional neural network. YT and SfZ critically revised drafted manuscript. All authors read and approved the final version of the manuscript. References F. Bray, J. Ferlay, I. Soerjomataram et al., "Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries", CA: a cancer journal for clinicians, vol. 68, no. 6, pp. 394–424, 2018. R. J. Kurman, "The Bethesda system for reporting cervical/vaginal cytologic diagnoses: definitions, criteria, and explanatory notes for terminology and specimen adequacy", Springer Science & Business Media, 2012. E. Jangam, A.A.D. Barreto, C.S.R. & Annavarapu, "Automatic detection of COVID-19 from chest CT scan and chest X-Rays images using deep learning, transfer learning and stacking.", Applied Intelligence , vol. 52, pp. 2243–2259, 2022. D. J. Chute, H. Lim and C. S. Kong, "BD focalpoint slide profiler performance with atypical glandular cells on SurePath Papanicolaou smears", Cancer cytopathology, vol. 118, no. 2, pp. 68–74, 2010. E. Bengtsson and P. Malm, "Screening for cervical cancer using automated analysis of PAP-smears", Computational and mathematical methods in medicine , 2014. W. William, A. Ware, A. H. Basaza-Ejiri, et al., "A review of image analysis and machine learning techniques for automated cervical cancer screening from pap-smear images", Computer methods and programs in biomedicine , vol. 164, pp. 15–22, 2018. S. U. Akram, J. Kannala, L. Eklund, et al., "Cell segmentation proposal network for microscopy image analysis", Deep Learning and Data Labeling for Medical Applications . Springer, Cham, pp. 21–29, 2016. P. Wang, L. Wang, Y. Li, et al., "Automatic cell nuclei segmentation and classification of cervical Pap smear images", Biomedical Signal Processing and Control, vol. 48, pp. 93–103, 2019. A. Ghoneim, G. Muhammad and M. S. Hossain, "Cervical cancer classification using convolutional neural networks and extreme learning machines", Future Generation Computer Systems, vol. 102, pp. 643–649, 2020. M. Ghasemi, M. Kelarestaghi, F. Eshghi, et al., "D 3 FC: deep feature-extractor discriminative dictionary-learning fuzzy classifier for medical imaging", Applied Intelligence, 2021. https://doi.org/10.1007/s10489-021-02781-w M. Q. Xu, W. X. Zeng, Y. H. Sun, et al., "Cervical cytology intelligent diagnosis based on object detection technology", in Processings of 1st Conference on Medical Imaging with Deep Learning (MIDL) , 2018, Amsterdam, The Netherlands. Y. Xiang, W. X. Sun, C. L. Pan, et al., "A novel automation-assisted cervical cancer reading method based on convolutional neural network", Biocybernetics and Biomedical Engineering, vol. 40, no. 2, pp. 611–623, 2020. Z. Zhuang, "Recognition of Cervical Cancer Cells Based on Improved ResNet Network", Beijing Jiaotong University, 2019. L. Liu, W. Ouyang, X. Wang, et al., "Deep Learning for Generic Object Detection: A Survey", International Journal of Computer Vision, vol. 128, no. 2, pp. 261–318, 2020. S. Ren, K. He, R. Girshick, et al., "Faster r-cnn: Towards real-time object detection with region proposal networks", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137–1149, 2017. K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition", arXiv preprint arXiv:1409.1556, 2014. K. He, X. Zhang, S. Ren, et al., "Deep residual learning for image recognition", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2016, Los Alamitos, USA. G. Huang, Z. Liu, L. V. Der Maaten, et al., "Densely Connected Convolutional Networks", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2017,Honolulu, HI, USA Z. Li, C. Peng, G. Yu, et al., "DetNet: Design Backbone for Object Detection", in European conference on computer vision, 2018, Munich, Germany. A. Geiger, P. Lenz and R. Urtasun, "Are we ready for autonomous driving? The KITTI vision benchmark suite", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2012. S. Xie, R. Girshick, P. Dollár, et al., "Aggregated residual transformations for deep neural networks", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pp. 1492–1500, 2017. J. Redmon and A. Farhadi, "YOLO9000: better, faster, stronger", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2017, Los Alamitos, CA: IEEE Computer Society Press. J. Pang, K. Chen, J. Shi, et al., "Libra r-cnn: Towards balanced learning for object detection", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2019,Los Alamitos, CA: IEEE Computer Society Press. L. D. Nguyen, D. Lin, Z. Lin, et al., "Deep CNNs for microscopic image classification by exploiting transfer learning and feature concatenation", in IEEE International Symposium on Circuits and Systems, 2018, Piscataway. C. Szegedy, V. Vanhoucke, S. Ioffe, et al., "Rethinking the Inception Architecture for Computer Vision", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2016, Los Alamitos, USA. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Major revision 06 May, 2022 Reviews received at journal 29 Apr, 2022 Reviewers agreed at journal 29 Apr, 2022 Reviews received at journal 29 Apr, 2022 Reviewers agreed at journal 28 Apr, 2022 Reviewers invited by journal 28 Apr, 2022 Editor assigned by journal 25 Apr, 2022 Editor invited by journal 25 Apr, 2022 Submission checks completed at journal 25 Apr, 2022 First submitted to journal 24 Mar, 2022 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About In Review Editorial Policies Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-1485286","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":101125031,"identity":"a0487725-fa66-4d8f-9c44-877f6d5690d8","order_by":0,"name":"Wei Wang","email":"","orcid":"","institution":"Key laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Department of Gynecologic Oncology, Peking University Cancer Hospital \u0026 Institute, Beijing,100142","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Wang","suffix":""},{"id":101125032,"identity":"5adf9136-3a94-471f-8ccd-1489141a2e67","order_by":1,"name":"Yun Tian","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxElEQVRIiWNgGAWjYDACZgglZwCm2IjQwQPVYkyCFiiduIFoLfbszM8e87bdSd8u3WPA8KHsMAP/7AZCDmMzN+Zte5a7c84ZA8YZ5w4zSNw5QEgLg5k0b9vh3A03cgyYgQwGA4kEQlrYv4G0pBuAtPwlTgsP2JYEsBZGorQc5imTnHPusOHOGWkFB3vOpfNI3CCghb3/+DaJN2WH5c0lkjc++FFmLcc/g4AWEGCCRc4BBkRE4QeMP4hSNgpGwSgYBSMWAACRxzvsnB/tWQAAAABJRU5ErkJggg==","orcid":"","institution":"School of Artificial Intelligence, Beijing Normal University, Beijing, 100875","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Yun","middleName":"","lastName":"Tian","suffix":""},{"id":101125034,"identity":"3a675242-dc72-4637-80ac-1ee69434a78a","order_by":2,"name":"Yang Xu","email":"","orcid":"","institution":"School of Artificial Intelligence, Beijing Normal University, Beijing, 100875","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Xu","suffix":""},{"id":101125035,"identity":"bf15011d-754c-497d-b75e-6ff743449dd8","order_by":3,"name":"Xiao-xuan Zhang","email":"","orcid":"","institution":"School of Artificial Intelligence, Beijing Normal University, Beijing, 100875","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiao-xuan","middleName":"","lastName":"Zhang","suffix":""},{"id":101125037,"identity":"bb3faf5a-ed90-4324-a669-6d526aefdce7","order_by":4,"name":"Yan-song Li","email":"","orcid":"","institution":"School of Artificial Intelligence, Beijing Normal University, Beijing, 100875","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Yan-song","middleName":"","lastName":"Li","suffix":""},{"id":101125039,"identity":"50016d69-ccc9-4ef8-921b-88def9b1b326","order_by":5,"name":"Shi-feng Zhao","email":"","orcid":"","institution":"School of Artificial Intelligence, Beijing Normal University, Beijing, 100875","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Shi-feng","middleName":"","lastName":"Zhao","suffix":""}],"badges":[],"createdAt":"2022-03-24 11:44:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-1485286/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-1485286/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":20983020,"identity":"ec8c0499-7979-4e9d-93bf-c4a0f5fbf600","added_by":"auto","created_at":"2022-05-02 13:35:27","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":143558,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverall flow of the proposed 3cDe-Net\u003c/strong\u003e\u003c/p\u003e\u003cp\u003e\u003cbr\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/6360b89b47b001486dbf9387.png"},{"id":20982854,"identity":"dce717df-c762-41ec-9fb7-28af3edcbb8f","added_by":"auto","created_at":"2022-05-02 13:30:27","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":63250,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNetwork structure of the proposed DC-ResNet\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/7cceb19231c2700e53e72113.png"},{"id":20983017,"identity":"d6a7667c-d2c7-4dc5-8d23-3dec85a1da09","added_by":"auto","created_at":"2022-05-02 13:35:27","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":53714,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGroup convolution\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/0f4eb0b56fe657cf1b761f2b.png"},{"id":20982852,"identity":"32333b64-b9b0-4f49-a7cd-5e7bd0dcb01f","added_by":"auto","created_at":"2022-05-02 13:30:27","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":45801,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eResidual group convolution block\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/f38357d11c8427be880105bc.png"},{"id":20983369,"identity":"83ae7f4a-971b-448a-9348-7613d6561753","added_by":"auto","created_at":"2022-05-02 13:40:27","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":45920,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eResidual dilated convolution block\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/2b88a9b6381f524cb8598df7.png"},{"id":20983019,"identity":"ebea858e-8787-4085-922c-5f2135323316","added_by":"auto","created_at":"2022-05-02 13:35:27","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":3451,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIoU calculation diagram\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/591707d030f9789bb9452541.png"},{"id":20982855,"identity":"9fab97d9-2b9c-48b9-a66f-cf40ad428b92","added_by":"auto","created_at":"2022-05-02 13:30:27","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":236998,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExample images of two datasets\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/89977e7970897b2e119abbdb.png"},{"id":20982858,"identity":"c9784a3e-f9f2-426f-93e4-f8cbc998cca8","added_by":"auto","created_at":"2022-05-02 13:30:27","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":352600,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExamples of DC-ResNet correct recognition results\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/ccaeed76f8189287ad23160d.png"},{"id":20982856,"identity":"a93efc95-f871-4ffc-a7fb-c7fc7ebf88cc","added_by":"auto","created_at":"2022-05-02 13:30:27","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":576681,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDetection examples of our 3cDe-Net\u003c/strong\u003e\u003c/p\u003e\u003cp\u003e\u003cbr\u003e\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/0c343ec20493ddae6b04da6b.png"},{"id":20983370,"identity":"6d19eac7-d89f-4e57-8887-cbc7342f6604","added_by":"auto","created_at":"2022-05-02 13:40:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1299584,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-1485286/v1/a97157ce-5cc2-47ec-b60f-42ac84156a68.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"3cDe-Net： A Cervical Cancer Cell Detection Network Based on an Improved Backbone Network and Multiscale Feature Fusion","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCervical cancer is the fourth most common gynecological malignancy globally. In 2018, there were approximately 570,000 new cases and 310,000 deaths worldwide [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Traditional cervical cancer cell screening typically requires a pathologist to observe thousands of cells under a microscope and provide a report based on diagnostic criteria[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. This approach is time-consuming, labor-intensive, relies heavily on the doctor's experience and has a strong subjectivity[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eComputer-aided cervical cancer cell detection is likely to become common for clinical diagnosis to solve the above problems [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. According to whether the segmentation step is included in the analysis pipeline, the identification of cervical cancer cells can be divided into segmentation-based recognition and objection detection-based recognition. Segmentation-based recognition methods typically segment cells or cell components and then extract cell characteristics for cell classification [\u003cspan additionalcitationids=\"CR7 CR8 CR9\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Detection accuracy typically depends on the results of cell segmentation, which makes it difficult to accurately identify overlapping cells with blurred cytoplasmic boundaries. Clinical applications face more difficulties. Object-detection-based cervical cancer cell recognition has been a trend in recent years[\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] and use an object detection framework based on a convolutional neural network to obtain the classification and location of cancer cells. According to whether regional candidate boxes are generated, object detection methods can be divided into two categories: two-stage methods and one-stage methods[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eXiang et al.[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] proposed an automatic assisted cervical cell screening system based on YOLO-v3-net and designed a classifier to further distinguish the categories of hard samples. Zhuang et al.[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] designed a special backbone network for cervical cancer cells and applied it to the single shot multibox detector (SSD) framework. These deep learning-based cervical cancer cell detection algorithms are one-stage [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], and their detection accuracy is not high. Xu et al.[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] proposed a two-stage detection method and transplanted the Faster R-CNN[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] framework of natural images into cervical cancer cells. However, the difference between cervical cancer cells and natural images is not considered; as a result, the detection performance is weak.\u003c/p\u003e \u003cp\u003eIn this study, we propose a two-stage cervical cancer detection algorithm based on an improved backbone network and multiscale feature fusion. In the first stage, cervical cell features with different sizes are extracted via an improved backbone network, and feature extraction can be verified through classification experiments. In the second stage, the location information of cervical cancer cells is obtained via the detection network. The features with different scales are fused through the feature pyramid, and then, the adaptive anchors are located by K-means clustering. Also, a loss function that alleviates the imbalance of sample contribution is defined, which improves detection accuracy. The proposed approach processes and analyzes samples directly at the picture level rather than the cellular level, which is more efficient and meets clinical needs more effectively. By horizontal comparison of cells on the image, the category information and location information can be obtained concurrently.\u003c/p\u003e \u003cp\u003eThe primary contributions of this study are as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eTo address the problem of different cell sizes, we present a two-stage cervical cancer detection method based on multiscale feature fusion.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe introduce dilated convolution and group convolution in the backbone into a deep residual network. The proposed 3cDe-Net avoids upsampling operations and reduces the information loss of small cells on the feature map;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eDue to the characteristics of cervical cancer cells, K-means clustering is used to determine the anchor frame size and ratio that are more suitable for cervical cancer cells and can providing better prior knowledge. We also define a loss function to address the imbalance between negative and positive samples for cervical cancer cell detection.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"2. Related Works","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Backbone network\u003c/h2\u003e \u003cp\u003eCurrently, deep learning-based detection algorithms must typically identify the features of an input picture through the feature extraction network. The feature extraction network of the classification task is also known as the backbone network. VGGNet[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] is the backbone network of Faster R-CNN[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], and the structure is simple. ResNet uses a deeper network structure to extract more complex features[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Both networks are still relatively common backbone networks. In addition, DenseNet densely connects each layer in the network with other layers, as well as DetNet specifically designed for object detection[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eExisting backbone networks are primarily used to recognize natural images. However, in cervical cell images, the canceration of cells is a gradually changing process, which makes it difficult to distinguish normal cells from cancerous cells using traditional backbone networks. Also, the scale of cervical cells varies, and small cells are difficult to identify on deeper feature maps, which makes it more difficult to detect these small cells. Thus, we use group convolution to enhance the expression ability of the extracted features. Concurrently, to better distinguish normal cells from abnormal cells, a dilated convolution is used to improve the resolution of the generated map, as well as the classification accuracy of small cells. Furthermore, this convolution also reduces the number of calculations caused by upsampling during feature map fusion.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Object detection based on deep learning\u003c/h2\u003e \u003cp\u003eCompared with classification, detection has an additional location task. Therefore, based on the backbone network, a network called the detection head should be added to locate the object region proposal. Thus, the backbone network and the detection head together construct the detection network. According to whether the detection head contains the region proposal network, the detection networks can be divided into one- and two-stage methods. The one-stage detection methods do not generate the region proposals, and the location and category prediction are completed in one stage. However, the accuracy of these methods could be improved. The two-stage methods first perform pre-detection by generating regional proposals and then fine-tuning the location and classification, yielding high accuracies. Faster R-CNN [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] is a classic two-stage detection network and is effective for many natural image datasets[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], but the network is not suitable for cervical cells due to the anchors generated based on natural images. In addition, the small cells retain less information on the feature map, which affects the detection accuracy of these cells. Finally, the loss function of the network does not consider the unbalanced sample size. To solve these problems, we use feature pyramids to integrate deep and shallow feature maps, which integrates deep semantic features and shallow location information more effectively. Therefore, small cells can be detected, and adaptive anchor boxes can be generated. Furthermore, a new loss function is defined, and the detection accuracy of cervical cancer cells is improved.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Proposed Methodology","content":"\u003cp\u003eThe overall framework of the proposed network is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The backbone network uses a dilated convolution ResNet (DC-ResNet). The algorithm first performs pre-detection of the cancer cells through the region proposal network (RPN) and then obtains the results through the classification and regression network.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMultiple feature maps with different scales are generated from the backbone network. Feature fusion is performed via the pyramid network to obtain the final predicted feature map, and then, the feature map is fed to the region proposal network. The feature pyramid network (FPN) achieves feature fusion by upsampling to make the depth and shallow feature maps with the same dimension. The feature pyramid can transfer deep semantic features to shallow layers to supplement semantic information. As a result, high-resolution and strong semantic features are obtained, which are fit to detect small objects. The RPN first adaptively generates anchor boxes on the generated prediction feature maps and then selects and adjusts the anchor boxes to obtain better region proposals. Next, the proposals and feature maps are fed into the classification and regression network. Finally, cervical cancer cells are predicted and located.\u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Improved Backbone Network: DC-ResNet\u003c/h2\u003e \u003cp\u003eDC-ResNet derived from ResNet is an improved backbone network. In DC-ResNet, a dilated convolution and a group convolution are introduced, and the details of the network structure is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The input images are fed to the network successively through the convolutional layer (Conv), batch normalization layer (BN), ReLU activation function and the pooling layer (MaxPooling), and then, the feature maps are obtained via multiple convolution groups. The first three groups (blue) use residual grouping convolution, and the last two groups (yellow) use residual dilated convolution.\u003c/p\u003e \u003cp\u003eTo improve the feature expression ability of the network, the feature map is fed into the fully connected layer to obtain the score of the predicted category. Each fully connected layer is followed by a dropout layer to prevent overfitting.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eGrouped convolution was originally a training method[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] that was designed to solve hardware resource limitations. To obtain more distinguishable features of cervical cancer cells, we group convolutional divisions on the channels, and then, the results of each group are \u003cem\u003econcatenated\u003c/em\u003e. The hyperparameter problem is solved by grouping convolution. Thus, model accuracy is improved without increasing the number of parameters. The convolution operation is performed by multiple GPUs, and the calculation results are connected. The calculation process is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(c\\)\u003c/span\u003e\u003c/span\u003e is the dimension before convolution, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(d\\)\u003c/span\u003e\u003c/span\u003e is the dimension after convolution.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn this study, residual group convolution is introduced, which is the introduction of group convolution based on residual networks, and the details are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The left panel is the overall structure diagram, and the right panel is the detailed diagram of the grouped convolution.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eResidual hole convolution uses hole convolution on the residual network. The calculation formula of the output feature size of the hole convolution is as follows:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}y=\u0026lfloor;\\frac{x-k-\\left(k-1\\right)\\times \\left(d-1\\right)+2p}{s}\u0026rfloor;+1\\#（1）\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(y\\)\u003c/span\u003e\u003c/span\u003e is the size of the output feature map, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(x\\)\u003c/span\u003e\u003c/span\u003e is the size of the input image, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(k\\)\u003c/span\u003e\u003c/span\u003e is the size of the convolution kernel, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(d\\)\u003c/span\u003e\u003c/span\u003e is dilation, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(p\\)\u003c/span\u003e\u003c/span\u003e is padding, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(s\\)\u003c/span\u003e\u003c/span\u003e is stride. The feature maps before and after convolution have the same parameters.\u003c/p\u003e \u003cp\u003eThe residual dilated convolution structure is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. The overall structure is similar to the residual group convolution, but the dilated convolution is used instead of the group convolution. The size of the feature map after the dilated convolution is unchanged. The residual dilated convolution has two structures A and B. The difference is whether the residual branch contains added 1\u0026times;1 convolution.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDownsampling/upsampling operations are not required for feature fusion. Thus, the size of the image feature map is larger than that of the original image. Thus, small cells are more informative at each feature point, and more edge information is retained.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Generation of anchor box\u003c/h2\u003e \u003cp\u003eWhen RPN generates region proposals, anchor boxes with different sizes on different feature layers of FPNs can be generated. The size and ratio of the anchor boxes are typically set based on domain prior knowledge or datasets. However, in this study, the changes in the size and the ratio of cancer cells were large, which is a challenge for anchor box generation. Inspired by YOLO v2[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], the best \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(k\\)\u003c/span\u003e\u003c/span\u003e boxes can be obtained by K-means clustering. The locations of all cells are not certain. We measure whether the size and ratio of anchors by looking at all the target boxes as being at the origin position. When clustering is performed, the distances are calculated between predicting boxes and centers by Eq.\u0026nbsp;(2):\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}b\\left(box,center\\right)=1-IoU\\left(box,center\\right)\\#\\left(2\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere IoU represents the fitting degree of the two boxes and is the ratio of the area of the intersection and the union of the prediction box and the real box, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe deep feature maps are small, and the receptive field is large, which is good for large cells. The shallow feature maps are the opposite, which is more suitable for detecting small cells. Therefore, anchors with different sizes and proportions can be generated on each feature map.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Definition of the loss function\u003c/h2\u003e \u003cp\u003eThe detection network is a multitask learning model that must predict the classification and location of cervical cancer cells. Therefore, the loss function should be considered to include classification and regression, and the definition is as follows:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}L\\left(\\left\\{{P}_{i}\\right\\},\\left\\{{bbox}_{i}\\right\\}\\right)=\\frac{1}{{N}_{cls}}\\sum _{i}{L}_{cls}\\left({p}_{i},{p}_{i}^{*}\\right)+\\lambda \\frac{1}{{N}_{reg}}\\sum _{i}{P}_{i}^{*}{L}_{reg}\\left({bbox}_{i},{bbox}_{i}^{*}\\right)\\#\\left(3\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\sum _{i}{L}_{cls}\\left({p}_{i},{p}_{i}^{*}\\right)\\)\u003c/span\u003e\u003c/span\u003e is the classification loss, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({p}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the real category, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({p}_{i}^{*}\\)\u003c/span\u003e\u003c/span\u003e is the predicting category. The classification function is calculated using cross-entropy loss, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\lambda\\)\u003c/span\u003e\u003c/span\u003e is the weight that balances the two task losses. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\sum _{i}{P}_{i}^{*}{L}_{reg}\\left({bbox}_{i},{bbox}_{i}^{*}\\right)\\)\u003c/span\u003e\u003c/span\u003e represents regression loss and only calculates positive samples, not including negative samples. The \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({smooth}_{L1}\\)\u003c/span\u003e\u003c/span\u003e function is used to calculate the regression loss. The definitions are as follows:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{L}_{reg}\\left({bbox}_{i},{bbox}_{i}^{*}\\right)=\\sum _{i\\in x,y,w,h}{smooth}_{L1}\\left({bbox}_{i}-{bbox}_{i}^{*}\\right)\\#\\left(4\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{smooth}_{L1}\\left(x\\right)=\\left\\{\\begin{array}{cc}0.5{x}^{2}\u0026amp; if\\left|x\\right|\u0026lt;1\\\\ \\left|x\\right|-0.5\u0026amp; otherwise\\end{array}\\right.\\#\\left(5\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eSetting the weights \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\lambda\\)\u003c/span\u003e\u003c/span\u003e to balance classification and the positioning loss remains challenging. The parameter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\lambda\\)\u003c/span\u003e\u003c/span\u003e is usually set manually. However, when calculating the regression task of the position, the model will consider the samples with a regression loss greater than 1 more when λ increases because the loss of the regression task is unconstrained. Therefore, when designing the loss function, more consideration should be given to those samples with a loss of less than 1. Inspired by Pang et al.[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], the original \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({smooth}_{L1}\\)\u003c/span\u003e\u003c/span\u003e is replaced with \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({balanced}_{L1}\\)\u003c/span\u003e\u003c/span\u003e defined as follows:\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{balanced}_{L1}\\left(x\\right)=\\left\\{\\begin{array}{cc}\\frac{\\alpha }{b}\\left(b\\left|x\\right|+1\\right)\\text{ln}\\left(b\\left|x\\right|+1\\right)\u0026amp; if\\left|x\\right|\u0026lt;1\\\\ \\gamma \\left|x\\right|+C\u0026amp; otherwise\\end{array}\\right.\\#\\left(6\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere α is used to control the gradient change of samples with a loss of less than 1 and γ is used to adjust the upper limit of the error. By adjusting the above two parameters, it is possible to balance the gradient contribution of each sample.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Experiments And Results","content":"\u003cp\u003e4.1 Datasets and evaluation metrics\u003c/p\u003e\n\u003cp\u003eThe experimental data evaluated in this study are from the Tian-chi competition datasets (Data-T)\u003csup\u003e[1\u003c/sup\u003e] and Herlev\u003csup\u003e[2]\u003c/sup\u003e. Fig. 7 shows several images from the dataset. In the Tian-chi competition dataset, each cervical cell smear image contains multiple cervical cells, which can be used for classification and detection. In the Herlev image dataset, each image contains only one cervical cell, which can be used for only classification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData-T:\u003c/strong\u003e This dataset comes from the preliminary data of the Cervical Cancer Risk Diagnosis Intelligent Challenge and contains 800 images of the Thinprep cytologic test (TCT) labeled by a professional pathologist, including 500 positive images and 300 negative images. Positive pictures were used to label the location of abnormal squamous epithelial cells. Due to the large size of the original pathological image (each image is approximately 40835 \u0026times;42371 pixels), we divided the original pathological image into several images with 800\u0026times;800 pixels to facilitate processing. Thus, there were 6,627 abnormal squamous epithelial cells. A negative sample refers to an image that does not contain cervical cancer cells.\u003c/p\u003e\n\u003cp\u003eDataset samples only labeled the location of abnormal squamous epithelial cells. In this study, 6627 normal squamous epithelial cells were screened from 300 negative pictures using semisupervised learning methods, and together with 6627 positive samples, a total of 13254 sample images were described by a classification dataset of cervical cancer cells. Images were divided into training sets, validation sets and test sets according to a ratio of 8:1:1, and the ratio of positive to negative samples was 1:1. To train the model with good generalizability, we enhanced the sample images using operations including rotation transformation and flip transformation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHerlev:\u003c/strong\u003e This dataset contains images of cervical cancer cells that were collected by Herlev University Hospital in Denmark. There are 917 single-cell images with 200\u0026times;100 pixels, including 242 normal cells and 675 abnormal cells. This dataset has become the primary study dataset for the classification of cervical cancer cells.\u003c/p\u003e\n\u003cp\u003eDue to the small amount of data and the imbalance between the positive and negative samples, we first augmented the sample images by center rotation and translation operations. Normal cell images were rotated 20 times, and abnormal cell images were rotated 10 times. As a result, the number of positive and negative samples is approximately equal. Finally, the training set, verification set, and test set were divided by a ratio of 8:1:1.\u003c/p\u003e\n\u003cp\u003eIn the classification experiments, metrics including sensitivity, specificity, h-mean, f1 measure and accuracy were used to evaluate the performance of the proposed feature extraction network. Sensitivity represents the proportion of correct images in all predicted cancer cell images, and specificity represents the proportion of correct images in all predicted normal cells. The detection experiment was evaluated by the mean average precision (mAP).\u003c/p\u003e\n\u003cp\u003e4.2 Network parameters and implementation details\u003c/p\u003e\n\u003cp\u003eThe size of the input images of the backbone network is 224\u0026times;224\u0026times;3, and the details of the network structure and parameters of DC-ResNet are listed in Table 1.\u003c/p\u003e\n\u003cp\u003eExperiments were performed on a workstation with the Ubuntu16.04 operating system and an NVIDIA GeForce 2080Ti 12 GB GPU. While training the backbone network, the SGD optimization algorithm was used to optimize the model parameters. The batch size was set to 32. The learning rate of each layer was initially set to 0.01. After 50 epochs of training, the learning rate was reduced to 1/10 every 10 epochs. The momentum was set to 0.9, and training ended after 1000 epochs. While training the detection network, the SGD optimization algorithm was used to optimize the model parameters. The batch size was set to 6, the learning rate of each layer was initially set to 0.00125, and the learning rate was reduced to 1/10 after 16 epochs and 22 epochs. The momentum was set to 0.9, and the weight was decayed by 0.001.\u003c/p\u003e\u0026nbsp;\u003ctable border=\"1\" id=\"Tab1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eStructure and parameters of DC-ResNet\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003eImproved backbone: DC-ResNet\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(7\\times 7\\)\u003c/span\u003e\u003c/span\u003e, 64, stride 2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(3\\times 3\\)\u003c/span\u003e\u003c/span\u003e, max pool, stride 2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"3\"\u003e\n \u003cp\u003eResidual group convolution\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\left[\\begin{array}{cc}1\\times 1\u0026amp; 128\\\\ 3\\times 3\u0026amp; 128\\\\ 1\\times 1\u0026amp; 256\\end{array}\\right]\\times 3\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGroup 32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\left[\\begin{array}{cc}1\\times 1\u0026amp; 256\\\\ 3\\times 3\u0026amp; 256\\\\ 1\\times 1\u0026amp; 512\\end{array}\\right]\\times 4\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGroup 32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\left[\\begin{array}{cc}1\\times 1\u0026amp; 512\\\\ 3\\times 3\u0026amp; 512\\\\ 1\\times 1\u0026amp; 1024\\end{array}\\right]\\times 6\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGroup 32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eResidual dilated convolution\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(B：\\left[\\begin{array}{cc}1\\times 1\u0026amp; 1024\\\\ 3\\times 3\u0026amp; 256\\\\ 1\\times 1\u0026amp; 1024\\end{array}\\right]\\times 1\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(A:\\left[\\begin{array}{cc}1\\times 1\u0026amp; 1024\\\\ 3\\times 3\u0026amp; 256\\\\ 1\\times 1\u0026amp; 1024\\end{array}\\right]\\times 2\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003edilation 2\u003c/p\u003e\n \u003cp\u003estride 2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(B：\\left[\\begin{array}{cc}1\\times 1\u0026amp; 1024\\\\ 3\\times 3\u0026amp; 256\\\\ 1\\times 1\u0026amp; 1024\\end{array}\\right]\\times 1\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(A:\\left[\\begin{array}{cc}1\\times 1\u0026amp; 1024\\\\ 3\\times 3\u0026amp; 256\\\\ 1\\times 1\u0026amp; 1024\\end{array}\\right]\\times 2\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003edilation 2\u003c/p\u003e\n \u003cp\u003estride 2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003efc-1024\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003efc-256\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003efc-2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e4.3 Results and analysis\u003c/p\u003e\n\u003cp\u003eWith the Data-H dataset, we used ResNet-50 and ResNet-101 as baselines for comparison with the proposed DC-ResNet. Table 2 lists the quantitative comparison results and shows that the proposed backbone network (DC-ResNet) performs better than baselines in all other metrics except for sensitivity. However, specificity is more important than sensitivity in the detection of cervical cancer cells due to the majority of normal cervical cell samples.\u003c/p\u003e\n\u003cp\u003eThe proposed DC-ResNet has 59 convolutional layers, and ResNet-50 has 50 convolutional layers. To verify the validity of the network structure of the proposed DC-ResNet, we compared DC-ResNet with ResNet-101 with 101 convolutional layers. Table 2 shows that DC-ResNet outperforms the other models. Also, all evaluation metrics, except for sensitivity, of the ResNet-101 network are lower than those of ResNet-50, which may be due to a limited number of datasets. Although no fittings of these complex models occurred, they do not necessarily produce better results.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;2 Quantitative comparison on the Data-H dataset\u003c/p\u003e\n\u003ctable border=\"1\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"18.411552346570396%\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.52346570397112%\"\u003e\n \u003cp\u003e\u003cstrong\u003eH-means\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.898916967509026%\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.425992779783392%\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"18.411552346570396%\"\u003e\n \u003cp\u003eResNet-50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.52346570397112%\"\u003e\n \u003cp\u003e96.82%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e96.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e96.98%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.898916967509026%\"\u003e\n \u003cp\u003e96.82%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.425992779783392%\"\u003e\n \u003cp\u003e96.83%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"18.411552346570396%\"\u003e\n \u003cp\u003eResNet-101\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.52346570397112%\"\u003e\n \u003cp\u003e96.75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e\u003cstrong\u003e97.12%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e96.37%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.898916967509026%\"\u003e\n \u003cp\u003e96.76%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.425992779783392%\"\u003e\n \u003cp\u003e96.75%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"18.411552346570396%\"\u003e\n \u003cp\u003eDC-ResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.52346570397112%\"\u003e\n \u003cp\u003e\u003cstrong\u003e97.11%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e95.92%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.870036101083034%\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.34%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.898916967509026%\"\u003e\n \u003cp\u003e\u003cstrong\u003e97.09%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.425992779783392%\"\u003e\n \u003cp\u003e\u003cstrong\u003e97.13%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eWith the Herlev dataset, we used Inception-v3[25], ResNet-152[17] and a feature concatenation network[24] as baselines for comparison with the proposed DC-ResNet. The details of the quantitative comparison are listed in Table 3, which shows that the proposed DC-ResNet achieves the highest classification accuracy by nearly 4%. Due to the small amount of data in the Herlev dataset, the 5-fold cross-validation method was used to verify the proposed network. Results indicate that the proposed DC-ResNet is superior to the baselines in terms of accuracy and exhibits better stability. The partial recognition results of DC-ResNet on the two datasets are shown in Fig. 8.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;3 Quantitative comparison on the Data-H dataset\u003c/p\u003e\n\u003ctable border=\"1\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"49.696969696969695%\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"50.303030303030305%\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"49.696969696969695%\"\u003e\n \u003cp\u003eInception-v3\u003csup\u003e[25]\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"50.303030303030305%\"\u003e\n \u003cp\u003e89.66\u0026plusmn;1.89%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"49.696969696969695%\"\u003e\n \u003cp\u003eResnet-152\u003csup\u003e[17]\u003c/sup\u003e\u003cstrong\u003e\u003csup\u003e\u0026nbsp;\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"50.303030303030305%\"\u003e\n \u003cp\u003e90.87\u0026plusmn;1.48%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"49.696969696969695%\"\u003e\n \u003cp\u003eFeature concatenation\u003csup\u003e[24]\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"50.303030303030305%\"\u003e\n \u003cp\u003e92.63\u0026plusmn;1.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"49.696969696969695%\"\u003e\n \u003cp\u003eDC-ResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"50.303030303030305%\"\u003e\n \u003cp\u003e\u003cstrong\u003e96.7%\u0026plusmn;1.1%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTo verify the effectiveness of DC-ResNet in the detection of cervical cancer cells, we performed a comparative experiment with ResNet-50 and ResNet-101 on mAP indicators, and the results are listed in Table 4. The mAP of ResNet-101 is not higher than that of ResNet-50, which shows that simply increasing the complexity of the model may lead to model overfitting with small datasets. Concurrently, the mAPs of DC-ResNet-50 are at least 1% higher than those of ResNet-50, which shows that the improved performance of the proposed model is not due to the increase in the number of network layers but rather the effect of changes in the network structure.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;4 mAP results of different backbone networks\u003c/p\u003e\n\u003ctable border=\"1\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eBackbone\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\[email protected]\u003c/strong\u003e\u003cstrong\u003e（\u003c/strong\u003e\u003cstrong\u003e%\u003c/strong\u003e\u003cstrong\u003e）\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\[email protected]\u003c/strong\u003e\u003cstrong\u003e（\u003c/strong\u003e\u003cstrong\u003e%\u003c/strong\u003e\u003cstrong\u003e）\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResNet-50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e45.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e26.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResNet-101\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e45.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e25.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eDC-ResNet\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e46.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e26.5\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eWe also analyzed the influence of the number of feature maps of DC-ResNet. F[1,2,3,4], F[1,2,3,5], F[1,2,4,5] and F[1,2,3,4,5] represent the feature fusion layers of the PFN of the corresponding feature layers in Fig. 1. The detection effect of feature fusion with different combinations is shown in Table 5. The worst detection effect is 46% when the input feature map is numbered [1,2,3,4]. Although the number of feature maps is reduced, the\u0026nbsp;mAP\u0026nbsp;is still 5% higher than ResNets. This result indicates that the superior performance of the proposed model is due to the increase in the number of feature layers and the structure of the proposed DC-ResNet itself.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;5 Results of FPN fusion of different layer feature maps\u003c/p\u003e\n\u003ctable border=\"1\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"66.9250645994832%\"\u003e\n \u003cp\u003e\u003cstrong\u003eThe number of fusion feature map\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.07493540051679%\"\u003e\n \u003cp\u003e\u003cstrong\[email protected] (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"66.9250645994832%\"\u003e\n \u003cp\u003e1，2，3，4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.07493540051679%\"\u003e\n \u003cp\u003e46.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"66.9250645994832%\"\u003e\n \u003cp\u003e1，2，3，5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.07493540051679%\"\u003e\n \u003cp\u003e46.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"66.9250645994832%\"\u003e\n \u003cp\u003e1，2，4，5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.07493540051679%\"\u003e\n \u003cp\u003e46.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"66.9250645994832%\"\u003e\n \u003cp\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003cstrong\u003e，\u003c/strong\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003cstrong\u003e，\u003c/strong\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003cstrong\u003e，\u003c/strong\u003e\u003cstrong\u003e4\u003c/strong\u003e\u003cstrong\u003e，\u003c/strong\u003e\u003cstrong\u003e5\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.07493540051679%\"\u003e\n \u003cp\u003e\u003cstrong\u003e46.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 6 shows the mAPs of the detection network with different improvement measures. The performance of the proposed 3cDe-Net is shown to be the best, with a maximum mAP of 50.4%. The reasons for the performance improvement of the proposed model are the anchor obtained by K-means clustering, the improved loss function, and DC-ResNet replacing the backbone network. The mAP metric increased by 3% due to the improvements in the detection network and the backbone network. The average IoU increased by 5% due to the anchor size generated by K-means clustering, and the mAP increased by 0.8% by introducing the generated anchor ratio into the detection network.\u003c/p\u003e\n\u003cp\u003eIn the improved\u0026nbsp;, the mAP increased by 1.2% with\u0026nbsp;\u0026nbsp;and\u0026nbsp;. By replacing the backbone network with DC-ResNet from the original ResNet-50, the mAP increased by 1.1%. Several detection results of our 3cDe-Net are shown in Fig. 9.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;6 Detection results of our 3cDe-Net\u003c/p\u003e\n\u003ctable border=\"1\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.56641604010025%\"\u003e\n \u003cp\u003eImproved anchor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.804511278195488%\"\u003e\n \u003cp\u003eImproved Loss\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"22.05513784461153%\"\u003e\n \u003cp\u003eDC-ResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.573934837092732%\"\u003e\n \u003cp\[email protected]\u0026nbsp;（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.56641604010025%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.804511278195488%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.05513784461153%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.573934837092732%\"\u003e\n \u003cp\u003e47.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.56641604010025%\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.804511278195488%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.05513784461153%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.573934837092732%\"\u003e\n \u003cp\u003e48.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.56641604010025%\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.804511278195488%\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.05513784461153%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.573934837092732%\"\u003e\n \u003cp\u003e49.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.56641604010025%\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.804511278195488%\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.05513784461153%\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.573934837092732%\"\u003e\n \u003cp\u003e50.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cdiv id=\"ftn1\"\u003e\n \u003cp\u003e1\u0026nbsp;https://tianchi.aliyun.com/competition/entrance/231757/introduction\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"ftn2\"\u003e\n \u003cp\u003e2 http://mde-lab.aegean.gr/index.php/downloads\u003c/p\u003e\n\u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eBased on the characteristics of cervical cancer cells, we proposed a multiscale feature fusion-based cervical cancer cell detection network based on an improved the backbone network, which contains a backbone network and a detection head. The backbone network was improved by group convolution and dilated convolution to optimize feature extraction, which increases the network receptive field while effectively downsampling. As a result, small cell features are more convenient to express. The detection head based on Faster R-CNN was improved, and different anchor boxes could be automatically set for different target sizes. Also, targets with different sizes were able to be predicted on different depths of feature layers. To minimize regression loss, a new balanced L1-based loss function was developed to reduce unbalanced sample contribution losses. Experiments with the Data-T and Herlev datasets show that the proposed model outperforms existing methods and achieves a mAP of 50.4%. The next step in this field of research is to identify the type and stage of cancer cells based on the identification of cervical cancer cells.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe confirm that all methods were carried out in accordance with relevant guidelines and regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets analysed during the current study are available in the Tian-chi Initiative and Herlev. https://tianchi.aliyun.com/competition/entrance/231757/introduction\u003c/p\u003e\n\u003cp\u003ehttp://mde-lab.aegean.gr/index.php/downloads\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that there is no conflict of interest regarding the publication of this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work has been partially supported by the National Natural Science Foundation of China (No. 62172047 and No. 61802020) and the Major Program of National Natural Science Foundation of China (No. 72091511).\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to express appreciation to the anonymous reviewers and editors for their helpful comments that improved the paper, and also would like to thank American Journal Experts for providing linguistic assistance during the preparation of this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; information\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e1\u003c/sup\u003eKey laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Department of Gynecologic Oncology, Peking University Cancer Hospital \u0026amp; Institute, Beijing,100142, China;\u003csup\u003e\u0026nbsp;\u003c/sup\u003e\u003csup\u003e2\u0026nbsp;\u003c/sup\u003eSchool of Artificial Intelligence, Beijing Normal University, Beijing, 100875, China.\u003c/p\u003e\n\u003cp\u003eCorrespondence should be addressed to\u0026nbsp;Yun Tian;\u0026nbsp;[email protected]\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWW, XxZ, and XY conceived the study and wrote the manuscript. XY, XxZ, and YsL performed training of convolutional neural network. YT and SfZ critically revised drafted manuscript. All authors read and approved the final version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eF. Bray, J. Ferlay, I. Soerjomataram et al., \"Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries\", CA: a cancer journal for clinicians, vol.\u0026nbsp;68, no. 6, pp.\u0026nbsp;394\u0026ndash;424, 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR. J. Kurman, \"The Bethesda system for reporting cervical/vaginal cytologic diagnoses: definitions, criteria, and explanatory notes for terminology and specimen adequacy\", Springer Science \u0026amp; Business Media, 2012.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eE. Jangam, A.A.D. Barreto, C.S.R. \u0026amp; Annavarapu, \"Automatic detection of COVID-19 from chest CT scan and chest X-Rays images using deep learning, transfer learning and stacking.\", \u003cem\u003eApplied Intelligence\u003c/em\u003e, vol.\u0026nbsp;52, pp.\u0026nbsp;2243\u0026ndash;2259, 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eD. J. Chute, H. Lim and C. S. Kong, \"BD focalpoint slide profiler performance with atypical glandular cells on SurePath Papanicolaou smears\", Cancer cytopathology, vol.\u0026nbsp;118, no. 2, pp.\u0026nbsp;68\u0026ndash;74, 2010.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eE. Bengtsson and P. Malm, \"Screening for cervical cancer using automated analysis of PAP-smears\", \u003cem\u003eComputational and mathematical methods in medicine\u003c/em\u003e, 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW. William, A. Ware, A. H. Basaza-Ejiri, et al., \"A review of image analysis and machine learning techniques for automated cervical cancer screening from pap-smear images\", \u003cem\u003eComputer methods and programs in biomedicine\u003c/em\u003e, vol.\u0026nbsp;164, pp.\u0026nbsp;15\u0026ndash;22, 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. U. Akram, J. Kannala, L. Eklund, et al., \"Cell segmentation proposal network for microscopy image analysis\", \u003cem\u003eDeep Learning and Data Labeling for Medical Applications\u003c/em\u003e. Springer, Cham, pp.\u0026nbsp;21\u0026ndash;29, 2016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eP. Wang, L. Wang, Y. Li, et al., \"Automatic cell nuclei segmentation and classification of cervical Pap smear images\", Biomedical Signal Processing and Control, vol.\u0026nbsp;48, pp.\u0026nbsp;93\u0026ndash;103, 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. Ghoneim, G. Muhammad and M. S. Hossain, \"Cervical cancer classification using convolutional neural networks and extreme learning machines\", Future Generation Computer Systems, vol.\u0026nbsp;102, pp.\u0026nbsp;643\u0026ndash;649, 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Ghasemi, M. Kelarestaghi, F. Eshghi, et al., \"D\u003csup\u003e3\u003c/sup\u003eFC: deep feature-extractor discriminative dictionary-learning fuzzy classifier for medical imaging\", Applied Intelligence, 2021. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10489-021-02781-w\u003c/span\u003e\u003cspan address=\"10.1007/s10489-021-02781-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Q. Xu, W. X. Zeng, Y. H. Sun, et al., \"Cervical cytology intelligent diagnosis based on object detection technology\", \u003cem\u003ein Processings of 1st Conference on Medical Imaging with Deep Learning (MIDL)\u003c/em\u003e, 2018, Amsterdam, The Netherlands.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Xiang, W. X. Sun, C. L. Pan, et al., \"A novel automation-assisted cervical cancer reading method based on convolutional neural network\", Biocybernetics and Biomedical Engineering, vol.\u0026nbsp;40, no. 2, pp.\u0026nbsp;611\u0026ndash;623, 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Zhuang, \"Recognition of Cervical Cancer Cells Based on Improved ResNet Network\", Beijing Jiaotong University, 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL. Liu, W. Ouyang, X. Wang, et al., \"Deep Learning for Generic Object Detection: A Survey\", International Journal of Computer Vision, vol.\u0026nbsp;128, no. 2, pp.\u0026nbsp;261\u0026ndash;318, 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Ren, K. He, R. Girshick, et al., \"Faster r-cnn: Towards real-time object detection with region proposal networks\", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.\u0026nbsp;39, no. 6, pp.\u0026nbsp;1137\u0026ndash;1149, 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK. Simonyan and A. Zisserman, \"Very deep convolutional networks for large-scale image recognition\", arXiv preprint arXiv:1409.1556, 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK. He, X. Zhang, S. Ren, et al., \"Deep residual learning for image recognition\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2016, Los Alamitos, USA.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG. Huang, Z. Liu, L. V. Der Maaten, et al., \"Densely Connected Convolutional Networks\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2017,Honolulu, HI, USA\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Li, C. Peng, G. Yu, et al., \"DetNet: Design Backbone for Object Detection\", in European conference on computer vision, 2018, Munich, Germany.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. Geiger, P. Lenz and R. Urtasun, \"Are we ready for autonomous driving? The KITTI vision benchmark suite\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2012.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Xie, R. Girshick, P. Doll\u0026aacute;r, et al., \"Aggregated residual transformations for deep neural networks\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pp.\u0026nbsp;1492\u0026ndash;1500, 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Redmon and A. Farhadi, \"YOLO9000: better, faster, stronger\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2017, Los Alamitos, CA: IEEE Computer Society Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Pang, K. Chen, J. Shi, et al., \"Libra r-cnn: Towards balanced learning for object detection\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2019,Los Alamitos, CA: IEEE Computer Society Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL. D. Nguyen, D. Lin, Z. Lin, et al., \"Deep CNNs for microscopic image classification by exploiting transfer learning and feature concatenation\", in IEEE International Symposium on Circuits and Systems, 2018, Piscataway.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eC. Szegedy, V. Vanhoucke, S. Ioffe, et al., \"Rethinking the Inception Architecture for Computer Vision\", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2016, Los Alamitos, USA.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Cervical cancer detection, Feature fusion, Backbone network, Adaptive anchors， Loss function","lastPublishedDoi":"10.21203/rs.3.rs-1485286/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-1485286/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eCervical cancer cell detection is an essential means of cervical cancer screening. However, in thin-prep cytology test-based images, the detection accuracy of traditional computer-aided detection algorithms is typically low due to the overlap of cells with blurred cytoplasmic boundaries. Clinical applications face more difficulties.\u003c/p\u003e\u003ch2\u003eMethod\u003c/h2\u003e \u003cp\u003ewe propose a cervical cancer cell detection network (3cDe-Net) based on an improved backbone network and multiscale feature fusion, and the network consists of the backbone network and the detection head. In the backbone network, a dilated convolution and a group convolution are introduced to improve the resolution and model expression ability. In the detection head, multiscale features are obtained based on a feature pyramid fusion network to ensure accurate capture of small cells; and then, based on Faster R-CNN, adaptive anchors of cervical cancer cells are generated via unsupervised clustering; furthermore, a new balanced L1-based loss function is defined, which reduces the unbalanced sample contribution loss.\u003c/p\u003e\u003ch2\u003eResult\u003c/h2\u003e \u003cp\u003eFor two different datasets(the Data-T dataset and Herlev dataset), the baselines including ResNet-50, ResNet-101, Inception-v3, ResNet-152 and the feature concatenation network were used, and the final quantitative results showed the effectiveness of the proposed backbone network DC-ResNet. Furthermore, experiments with both datasets show that our detection network 3cDe-Net, based on the optimal anchor, the defined new loss function, and the DC-ResNet, outperforms existing methods and achieves a mAP of 50.4%. By horizontal comparison of cells on the image, the category information and location information can be obtained concurrently.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe proposed 3cDe-Net is an end-to-end network that can detect cancer cells on multicell pictures and can determine their locations. The model processes and analyzes samples directly at the picture level rather than the cellular level, which is more efficient and meets clinical needs more effectively.\u003c/p\u003e","manuscriptTitle":"3cDe-Net： A Cervical Cancer Cell Detection Network Based on an Improved Backbone Network and Multiscale Feature Fusion","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2022-05-02 13:30:25","doi":"10.21203/rs.3.rs-1485286/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revision","date":"2022-05-06T20:22:50+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2022-04-29T09:15:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"01f131a3-3ec2-4df1-867c-c0ba5728418d","date":"2022-04-29T08:55:08+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2022-04-29T08:48:28+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"8bd630a5-deaa-4646-908b-dd8eac027689","date":"2022-04-28T12:10:20+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2022-04-28T12:06:43+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2022-04-25T08:53:40+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2022-04-25T07:54:52+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2022-04-25T07:51:53+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2022-03-24T11:38:32+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b210e11e-b0bd-4c6b-88d0-83a228ac4d39","owner":[],"postedDate":"May 2nd, 2022","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2022-07-05T06:14:19+00:00","versionOfRecord":[],"versionCreatedAt":"2022-05-02 13:30:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-1485286","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-1485286","identity":"rs-1485286","version":["v1"]},"buildId":"pf3fE39SIOqb-0xH_OWvX","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}