CircleSnake: Instance Segmentation with Circle Representation

Circle representation has recently been introduced as a “medical imaging optimized" for more effective instance object detection on ball-shaped medical objects. With its superior performance detection, it is appealing to extend the circle segmentation. In this work, we propose CircleSnake, simple end-to-end contour deformation-based segmentation method Compared prevalent DeepSnake method, our contribution threefold: (1) We replace complicated bounding box octagon transformation with computation-free and consistent adaption segmenting objects; (2) fewer degrees of freedom (DoF = 2) compared 8), thus yielding robust better rotation consistency; (3) To best knowledge, proposed CircleSnake first deep pipeline proposal, circular convolution. The key innovation integrate graph convolution into an framework, enabled by representation. Glomeruli are used evaluate benchmarks. From results, increases average precision glomerular from 0.559 0.614. Dice score increased 0.804 0.849. code released: .