investigating Explainable Artificial intelligence For Mri-Based Classification of Dementia: A New Stability Criterion For Explainable Methods

Segmenting instances is a challenging task, especially for precise boundary segmentation. Modern methods always predict instance masks with imprecise boundaries due to the fact they do not pay enough attention to boundary information. To address this problem, we propose a conceptually simple yet effective attention-based module for better boundary segmentation, termed boundary-area enhanced module. in this model, we extend the thin boundary lines to boundary areas by broadening one line using its inner pixels within a distance. Boundary areas can be captured more explicitly and accurately, and then low-level features and high-level features are fused to predict instance masks better. The proposed boundary-area enhanced model yields significant improvements over the Mask R-CNN framework baseline on COCO and Cityscapes datasets. Extensive experiments show that our boundary-area enhanced module outperforms previous state-of-the-art models and can help other two-stage segmentation models to achieve better performance.