WNet: A study of a class-waveform attention-enabled deep learning method and its application to medical image segmentation.

Colorectal cancer is a common and fatal malignant tumor that has a high prevalence in both men and women. For this disease, early detection of the problem area is crucial. However, colorectal cancer usually has complex structures and textures and may contain irregular shapes, color variations, and blurred edges, making it difficult for the model to learn features accurately. To alleviate these problems, this study investigates the design of a class-waveform attention mechanism and utilizes this attention to guide the medical image segmentation model proposed in this study, which effectively alleviates the problem of accurate localization of irregular lesion regions by the model. In addition, on the basis of this attention mechanism, this study further designs an interlayer synchronization module using a channel separation-merging structure, through which the structural features of colorectal cancer are better captured, and the learning ability of the model is improved. To improve the effective fusion of features by the model, a fusion module is designed in this study using the gating concept and embedded into the up-sampling stage. To demonstrate the effectiveness of the proposed WNet model, this study evaluates the segmentation capability of the network on six sub-datasets from the EBHI-Seg dataset. Extensive experimental results show that WNet achieves the state-of-the-art segmentation performance in terms of mDice, mIoU, mPrecision and mRecall metrics. Our code will be available at https://github.com/DeadlyCodeGod/WNet.