A Diversified Supervised based U-shape Colorectal Lesion Segmentor with Meaningful Feature Supplement and Multi-Level Residual Attention Mechanism

Colorectal cancer is a commonly diagnosed cancer of digestive system. Automatic and accurate segmentation of colorectal tumors from medical images (e.g., CT) has great significance for diagnosis, staging and treatment planning. However, the blurred boundary of tumors, as well as variability of their location and shape, make most traditional methods ineffectual. In this paper, we propose a diversified supervised U-shape CNN colorectal lesion segmentor (DSUCLS) to overcome this challenge. Our model mainly contains three key components: 1) the weakly supervised transfer learning module for supplementing generic features, where the irrelevant ones are filtered out by extra convolutional layers and image-level label, 2) an encoder-decoder structure based on U-shape architecture for learning specific pathological representation from medical images, 3) the multilevel supervised attention module incorporated into decoder path for producing coarse-to-fine guidance and guaranteeing finer attention map. 4), the pre-processing and post-processing strategies are applied to further improve segmentation performance. The experimental results illustrate that the proposed model outperforms other state-of-the-art techniques for colorectal lesion segmentation on CT images, achieving Dice scores of 0.733 and dramatically decreasing Hausdorff distance to 17.62.