Cervical precancerous lesions classification using pre-trained densely connected convolutional networks with colposcopy images

摘要

Colposcopy is currently a common medical technique for preventing cervical cancer. However, with the increase of the workload, screening by artificial vision has the problems of misdiagnosis and low diagnostic efficiency. Based on transfer learning, pre-trained densely connected convolutional networks are used to propose a computer-aided-diagnosis (CAD) method for automatic classification of cervical precancerous. The proposed method is applied to determine CIN2 or higher-level lesions in cervical images. In the present work, image data are initially prepossessing with ROI extraction and data augmentation. Then, parameters of all layers are fine-tuning with pre-trained DenseNet convolutional neural networks from two datasets (ImageNet and Kaggle). The impact of different training strategies on the model performance with limited training data is analyzed, including random initialization (RI) training from scratch, fine-tuning (FT) pre-trained model, different size of training data and K-fold cross-validation. Experimental results show that our method (FT) achieves an accuracy of 73.08% (AUC approximate to 0.75) in 600 test images. Compared with previous related work and clinicians, the performance of our approach can effective diagnosis CIN2+ and comparable with a senior physician, which proves the feasibility and promising of the proposed computer-aided diagnostic method. (C) 2019 Elsevier Ltd. All rights reserved.