HSIL Colposcopy Image Segmentation Using Improved U-Net

摘要

The screening and diagnosis of cervical lesions can effectively reduce the risk of cervical cancer. At present, there are few research algorithms and models for colposcopy in cervical lesion image segmentation. If image about colposcopy can be directly used to automatically segment the precancerous lesion area of cervical cancer, it will help doctors to make a faster and more accurate diagnosis, which has important clinical significance. To resolve this conundrum, this paper proposes an improved U-Net network image segmentation method for cervical squamous intraepithelia[1] lesions. Firstly, a sub-coding module is embedded in each layer of the U-Net encoder structure to improve its feature extraction capability. Secondly, dense connection is added to the lowest level convolution operation to solve the problem that the gradient disappears as the network deepens. Compared with U-Net, FCN and SegNet, the precision, Dice coefficient, IOU and Recall evaluation indexes of the model were better in this study, which effectively realized the image segmentation of cervical squamous intraepithelial lesions.