Research and Implementation of Colorectal Cancer Tumour Segmentation Algorithm based on DRA-UNet

Abstract

Fully convolutional neural networks often suffer from information loss and suboptimal accuracy in medical image segmentation tasks. To mitigate these issues for rectal tumor analysis, we propose an improved U-Net architecture, designated as Deep Residual Attention U-Net (DRA-UNet). First, deep residual modules replace the original convolutional blocks. This substitution enhances feature extraction capabilities while mitigating network degradation. secondly, an attention mechanism is added between each jump connection of the U-Net to focus attention on the features useful for segmentation and suppress redundant features; finally, a DiceLoss loss function is used to alleviate the class imbalance problem. Experimental validation was performed on a colorectal cancer CT image dataset, with segmentation performance quantified using the Dice coefficient. The experimental results demonstrate that the proposed method achieves a segmentation accuracy of 91.38% for colorectal cancer tumors. This represents a significant improvement over existing models, with performance gains of 9.08%, 6.61%, and 4.73% compared to the FCN, U-Net, and Dense-UNet frameworks, respectively. These comparative outcomes confirm the effectiveness and superiority of our approach.