MIA-Net: Multi-information aggregation network combining transformers and convolutional feature learning for polyp segmentation

Abstract

Accurate polyp segmentation is of immense importance for the early diagnosis and treatment of colorectal cancer. However, polyp segmentation is a difficult task, and most current methods suffer from two challenges. First, individual polyps widely vary in shape, size, and location (intra-class inconsistency). Second, subject to conditions such as motion blur and light reflection, polyps and their surrounding background have a high degree of similarity (inter-class indistinction). To overcome intra-class inconsistency and inter-class indistinction, we propose a multi-information aggregation network (MIA-Net) combining transformer and convolutional features. We use the transformer encoder to extract powerful global features and better localize polyps with an advanced global contextual feature extraction module. This approach reduces the influence of intra-class inconsistency. In addition, we capture fine-grained local texture features using the convolutional encoder and aggregate them with high-level and low-level information extracted by the transformer. This rich feature information makes the model more sensitive to edge information and alleviates inter-class indistinction. We evaluated the new approach quantitatively and qualitatively on five datasets using six metrics. The experimental results revealed that MIA-Net has good fitting ability and strong generalization ability. In addition, MIA-Net significantly improved the accuracy of polyp segmentation and outperformed the current state-of-the-art algorithms.