Cervical lesion image enhancement based on conditional entropy generative adversarial network framework

Abstract

Early screening and diagnosis of cervical precancerous lesions are very important to prevent cervical cancer. High-quality colposcopy images will help doctors make faster and more accurate diagnoses. To tackle the problem of low image quality caused by complex interference during colposcopy operation, this paper proposed a conditional entropy generative adversarial networks framework for image enhancement. A decomposition network based on Retinex theory is constructed to obtain the reflection images of the low-quality images, then the conditional generative adversarial network is used as the enhancement network. The low-quality images and the decomposed reflection images are both input the enhancement network for training, and the conditional entropy distance is used as a part of the loss of the conditional generative adversarial network to alleviate the over-fitting problem during the training process. The test results show that compared with published methods, the proposed method of this paper can significantly improve the image quality, and can enhance the colposcopy image while retaining image details.

Keywords: Cervical cancer; Conditional Generative Adversarial Network; Conditional entropy distance; Medical image enhancement; Retinex.