Cervical lesion segmentation via transformer-based network with attention and boundary-aware modules

Abstract

Colposcopic diagnosis and directed biopsy is the foundation of cervical cancer screening. In the procedure of colposcopy, automatic segmentation of cervical lesion in colposcopic images can provide great assistance and convenience especially in underdeveloped region. However, the existing methods based on Convolutional Neural Networks only differentiate the abnormality from healthy tissue, which is hard to further subdivide the lesion. In this paper, a Transformer-based network TABNet is proposed which can precisely extract the cervical lesion and recognize the corresponding category of each lesion. Unlike the other CNN-based methods, a more powerful vision transformer is adopted as the encoder. Three effective modules in decoder are constructed to integrate the advance in attention mechanism and boundary-aware prior knowledge. Extensive experiments on a large clinical colposcopic image dataset show that TABNet outperforms the existing state-of-art methods and achieves great improvement. Compared with nnUNet, our proposed model improves the mean DSC by 7.74% and mean IoU by 8.51%, respectively.