Neural Network Based Polyp Image Segmentation

As the third most common malignant tumor and the second deadliest disease,most colorectal cancers develop from colon polyps.Polyp segmentation,which aims to understand medical images with polyps at the pixel level and segment the outline of the lesion area,plays an important role in assisting doctors in diagnosis.Traditional segmentation methods mainly identified low-level features to segment polyp regions.However,due to the lack of high-level semantic features,traditional polyp segmentation methods usually achieve lower-quality segmentation performance and poor generalization ability.With the development of deep learning in the field of image recognition,polyp segmentation has made impressing progress.However,this task still faces challenges of blurred boundaries,large polyp size variations,and inconsistent internal colors.In addition,the heavy computational burden also makes deep learning models unable to meet the real-time response of practical applications.There are three mainstream deep learning methods,including convolutional neural network,the Transformer and multi-layer perceptron.Convolutional neural networks have a strong ability to extract local details.But they cannot capture long-range dependencies at high resolution due to the inherent locality of convolution operations.The Transformer networks utilize self-attention mechanism to extract long-term context information,while suffering from heavy computing cost.Benefiting from simple structures,MLP-based models seem to be an alternative.To solve the problems of polyp segmentation,this paper combines the advantages of three neural networks and proposes two methods as follows:1)An efficient context-aware MLP-based paradigm for polyp segmentation is proposed.To overcome the problem of large polyp size variations,Cycle MLP encoder is used to extract features of different scales and expand the receptive field.To solve the problems of blurred polyp boundaries and different colors inside polyps,two novel MLP-based decoders are proposed,including the Multi-head Mixer Module and the Contextual Bridger Module.Experiments demonstrate that the proposed method achieves the best performance on 4 public benchmarks with only 16 M parameters.2)A real-time CNN and Transformer hybrid network for polyp segmentation is proposed.The Topformer is utilized as a lightweight backbone network to extract multi-scale features and a multi-step fusion decoder is proposed to extract robust features.Experiments demonstrate that the proposed method can achieve comparable performance with only 5M parameters,and the inference speed reaches 64 FPS.