Detection of cervical lesion region from colposcopic images based on feature reselection

Colposcopy is one of the important steps in the clinical screening of cervical intraepithelial neoplasia (CIN) and early cervical cancer. It directly affects the patient's diagnosis and treatment program. Therefore, it is widely used for cervical cancer screening. The present work proposes a cervical lesion detection net (CLDNet) model based on the deep convolutional neural network (CNN). The Squeeze-Excitation convolutional neural network (SE-CNN) employed to extract depth features of the whole image. SE module for feature recalibration. Moreover, the region proposal network (RPN) generated a proposal box of the region of interest (ROI). Finally, the region of interest classified and proposal box regression performed to locate the cervical lesion region. The Squeeze-Excitation (SE block) strengthened important features and suppress non-primary features, improve feature extraction ability, which is beneficial to feature classification and proposal box regression in the regions of interest. It is found that the average precision of the model extraction lesion region is 92.53 % and the average recall rate is 85.56 %, which can play a good role in the auxiliary diagnosis.