Cervical Lesions Classification Based on Pre-trained MobileNet Model

This paper aims to establish an intelligent diagnosis model of cervical cancer screening and to solve the shortcomings of the physician and traditional computer-aided diagnosis methods in the current. We propose a computer-aided diagnosis method based on transfer learning, which uses the pre-trained MobileNetV2 model to classify colposcopic images. Firstly, the data is augmented and normalized, and then the MobileNetV2 model pre-trained on ImageNet is used to realize the classification diagnosis of cervical lesions in colposcopic images. Finally, the diagnosis results are compared with those of colposcopic physicians. Experiments show that this method can effectively diagnose CIN2+ lesions with an accuracy rate of 75.00%, which is higher than the average level of diagnosis by colposcopy physicians. This method overcomes the shortcomings of physicians’ diagnoses to a certain extent. The efficiency of CIN2+ lesion classification for colposcopy images is superior to other mainstream models, which is greatly significant for the current cervical lesion screening.