Computer-Aided Polyp Image Segmentation Using A Residual Attention Convolutional Neural Network

With a promising application in Computer-Aided Diagnosis (CAD) pipelines, medical image segmentation is the most commonly studied subject in the field of deep learning-based medical imaging. Yet existing methods failed to achieve satisfactory performance in pixel-level segmentation, thanks in part to the class imbalance between foreground and background regions. In addition, the robustness and generalizability of segmentation models lacked due attention in previous research. In this paper, we presented a residual attention convolutional network for polyp segmentation, which took advantage of residual learning, multi-level feature fusion, and channel-wise and spatial attention for adaptive feature calibration. For evaluation of the competitiveness of the proposed algorithm, we performed extensive experiments using five publicly accessible polyp datasets. Experimental results showed that our method enjoyed consistent improvement over state-of-the-art algorithms. For the assessment of segmentation models’ robustness and generalizability, we further performed cross-dataset and challenging-scenario evaluations. Experiments verified that our network can well generalize across different datasets and was exceptionally robust against challenging images.