Cascade Transformer Encoded Boundary-Aware Multibranch Fusion Networks for Real-Time and Accurate Colonoscopic Lesion Segmentation

Automatic segmentation of colonoscopic intestinal lesions is essential for early diagnosis and treatment of colorectal cancers. Current deep learning-driven methods still get trapped in inaccurate colonoscopic lesion segmentation due to diverse sizes and irregular shapes of different types of polyps and adenomas, noise and artifacts, and illumination variations in colonoscopic video images. This work proposes a new deep learning model called cascade transformer encoded boundary-aware multibranch fusion networks for white-light and narrow-band colorectal lesion segmentation. Specifically, this architecture employs cascade transformers as its encoder to retain both global and local feature representation. It further introduces a boundary-aware multibranch fusion mechanism as a decoder that can enhance blurred lesion edges and extract salient features, and simultaneously suppress image noise and artifacts and illumination changes. Such a newly designed encoder-decoder architecture can preserve lesion appearance feature details while aggregating the semantic global cues at several different feature levels. Additionally, a hybrid spatial-frequency loss function is explored to adaptively concentrate on the loss of important frequency components due to the inherent bias of neural networks. We evaluated our method not only on an in-house database with four types of colorectal lesions with different pathological features, but also on four public databases, with the experimental results showing that our method outperforms state-of-the-art network models. In particular, it can improve the average dice similarity coefficient and intersection over union from (84.3%, 78.4%) to (87.0%, 80.5%).