Detect, Grow, Seg: A weakly supervision method for medical image segmentation based on bounding box

Abstract

Weakly supervised semantic segmentation based on bounding box has been fully developed in natural scenes, but in medical image scenes it often faces the difficulties of blurred edges, imaging noise, and mutual interference between multiple organs, which makes the methods suitable in natural scenes tend to show poor performance in medical images. In this paper, we propose a weakly supervised segmentation method for medical images based on Detection, Growth, and Segmentation (DGS). For weak labels in the form of bounding boxes, the detector YOLOV5 and GradCAM are introduced to obtain the heatmap representation, while the additional unlabeled data can be also utilized through the detector. The thresholded heatmap is used to generate seed points for adaptive regional growth so that better pseudo labels are obtained. Finally, the pseudo labels are used for supervised training on UNet to obtain end to end segmentation. Sufficient experiments show that, compared with other methods, the pseudo label generation and segmentation model formed by our method achieve the best performance, while the use of unlabeled data by the detector brings significant improvement for segmentation in low-label scenarios, which greatly reduce the time and demand of labeling.