Attention-based Mask R-CNN for Microvascular Segmentation

The normal functioning of human organs and tissues depends on the spatial organization, specialization, and interaction between cells. With a total of 37 trillion cells in the human body, determining their functions and relationships is highly challenging. The Vascular Coordinate Common Framework (VCCF) is widely used for cell mapping, but it is limited by a lack of knowledge about the microvascular system. Therefore, we propose a novel framework called Attention-based Mask R-CNN for Microvascular Segmentation to achieve automated segmentation of microvascular arrangements. In terms of structure, our model replaces the original fully connected layers with the Feature Pyramid Network (FPN) as the segmentation head network and integrates feature maps from different levels. To enhance performance, we introduce RoI (Region of Interest) feature alignment, employing techniques such as bilinear interpolation to accurately align features within the RoI region. Furthermore, to better focus on the regions of interest, we introduce an attention mechanism specifically designed for Human Vasculature, which improves the segmentation accuracy of critical areas. We train our model using 2D PAS-stained histological images of healthy human kidney tissue sections. Experimental results demonstrate the superior segmentation outcomes of our approach for microvascular structures, including capillaries, arterioles, and venules.