An improved classification diagnosis approach for cervical images based on deep neural networks

Abstract

In order to enhance the speed and performance of cervical diagnosis, we propose an improved Residual Network (ResNet) by combining pyramid convolution with depth-wise separable convolution to obtain the high-quality cervical classification. Since most of cervical images from patients are not in the center of colposcopy images, we devise the segmentation and extraction algorithm of the center movement of the region of interest (ROI), which will further enhance the classification performance. Extensive experiments indicate that our model can not only achieve the lightweight network model, but also fulfil the classification prediction, such as for three-classification of cervical lesions, the classification accuracy is as high as 91.29\(\%\), the precision is 89.70\(\%\), the sensitivity is 88.75\(\%\), the specificity is 94.98\(\%\), the rate of missed diagnosis is 11.25\(\%\) and the rate of misdiagnosis is 5.02\(\%\). Finally, after dividing the colposcopy images into four categories, it is shown that our results are still better than those obtained from many previous works as far as the cervical image classification is concerned. The current work can not only assist doctors to quickly diagnose cervical diseases, but also the classification performance can meet some clinical requirements in practice.