Cervical Precancerous Lesion Detection Based on Deep Learning of Colposcopy Images

Abstract

With the rapid development of deep learning, automatic lesion detection is widely used in clinical screening. In this paper, we make use of convolutional neural network (CNN) algorithm to help medical experts detect cervical precancerous lesion during the colposcopic screening, especially in the classification of cervical intraepithelial neoplasia (CIN). Firstly, the original image data is classified into six categories: normal, cervical cancer, mild (CIN1), moderate (CIN2), severe (CIN3) and cervicitis, which are further augmented to solve the problem of few samples of endoscopic images and non-uniformity for each category. Then, a CNN-based model is built and trained for the multi-classification of the six categories, we have added some optimization algorithms to this CNN model to make the training parameters more effective. For the test dataset, the accuracy of the proposed CNN model algorithm is 89.36%, and the area under the receiver operating characteristic (ROC) curve is 0.954. Among them, the accuracy is increased by 18%–32% compared with other traditional learning methods, which is 9%–20% higher than several commonly used deep learning models. At the same number of iterations, the time consumption of proposed algorithm is only one quarter of other deep learning models. Our study has demonstrated that cervical colposcopic image classification based on artificial intelligence has high clinical applicability, and can facilitate the early diagnosis of cervical cancer.