Segmentation of the cervical lesion region in colposcopic images based on deep learning

Hui Yu; Yinuo Fan; Huizhan Ma; Haifeng Zhang; Chengcheng Cao; Xuyao Yu; Jinglai Sun; Yuzhen Cao; Yuzhen Liu

doi:10.3389/fonc.2022.952847

Segmentation of the cervical lesion region in colposcopic images based on deep learning

Front Oncol. 2022 Aug 3:12:952847. doi: 10.3389/fonc.2022.952847. eCollection 2022.

Authors

Hui Yu^{1

2}, Yinuo Fan¹, Huizhan Ma², Haifeng Zhang³, Chengcheng Cao³, Xuyao Yu⁴, Jinglai Sun², Yuzhen Cao², Yuzhen Liu³

Affiliations

¹ Academy of Medical Engineering and Translational Medicine, Tianjin University, Tianjin, China.
² School of Precision Instrument and Optoelectronics Engineering, Tianjin University, Tianjin, China.
³ Obstetrics and Gynecology, Affiliated Hospital of Weifang Medical University, Weifang, China.
⁴ Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, China.

Abstract

Background: Colposcopy is an important method in the diagnosis of cervical lesions. However, experienced colposcopists are lacking at present, and the training cycle is long. Therefore, the artificial intelligence-based colposcopy-assisted examination has great prospects. In this paper, a cervical lesion segmentation model (CLS-Model) was proposed for cervical lesion region segmentation from colposcopic post-acetic-acid images and accurate segmentation results could provide a good foundation for further research on the classification of the lesion and the selection of biopsy site.

Methods: First, the improved Faster Region-convolutional neural network (R-CNN) was used to obtain the cervical region without interference from other tissues or instruments. Afterward, a deep convolutional neural network (CLS-Net) was proposed, which used EfficientNet-B3 to extract the features of the cervical region and used the redesigned atrous spatial pyramid pooling (ASPP) module according to the size of the lesion region and the feature map after subsampling to capture multiscale features. We also used cross-layer feature fusion to achieve fine segmentation of the lesion region. Finally, the segmentation result was mapped to the original image.

Results: Experiments showed that on 5455 LSIL+ (including cervical intraepithelial neoplasia and cervical cancer) colposcopic post-acetic-acid images, the accuracy, specificity, sensitivity, and dice coefficient of the proposed model were 93.04%, 96.00%, 74.78%, and 73.71%, respectively, which were all higher than those of the mainstream segmentation model.

Conclusion: The CLS-Model proposed in this paper has good performance in the segmentation of cervical lesions in colposcopic post-acetic-acid images and can better assist colposcopists in improving the diagnostic level.

Keywords: cervical lesion; colposcopic images; deep learning; feature extraction; image segmentation.