Cau-Net: Enhancing Medical Image Segmentation With Contour-Guided Attention for Accurate Stroke Prediction

Accurate segmentation of medical images is necessary for accurate diagnosis and treatment planning. The achievement of the precise boundaries in complex anatomical structures is difficult. In this paper, a novel hybrid segmentation model called Contour-Guided Attention U-Net (CAU-Net) is proposed by combining the traditional region-based active contour models with the advanced deep learning capabilities of Attention U-Net. CAU-Net employs an active contour model that generates an initial segmentation mask that provides the boundary awareness needed to steer the attention U-Net refinement. The use of attention mechanisms will be allowed with the selective features on the regions of interest identified by active contours in order to refine it to high accuracy of results in segmentation. It was tested on a Kaggle MRI dataset of medical images, including stroke prediction tasks, and performed better in boundary accuracy and overall segmentation quality than standalone Attention U-Net and active contour models. In stroke prediction, the CAU-Net achieved 98.3% segmentation accuracy and a Dice coefficient of 0.87 in detecting stroke lesions and delineating fine anatomical details. From the results, the hybrid approach significantly amplifies the contour adherence in addition to contextual sensitivity such that these are effective methods in such complex segmentation task especially those in medical applications.