Weakly supervised glottis segmentation on endoscopic images with point supervision

The ability to automatically segment anatomical targets on medical images is crucial for clinical diagnosis and interventional therapy. However, supervised learning methods often require a large number of pixel-wise labels that are difficult to obtain. This paper proposes a weakly supervised glottis segmentation (WSGS) method for training end-to-end neural networks using only point annotations as training labels. This method functions by iteratively generating pseudo-labels and training the segmentation network. An automatic seeded region growing (ASRG) algorithm is introduced to generate quality pseudo labels to diffuse point annotations based on network prediction and image features. Additionally, a novel loss function based on the structural similarity index measure (SSIM) is designed to enhance boundary segmentation. Using the trained network as its core, a glottis state monitor is developed to detect the motion behavior of the glottis and assist the anesthesiologist. Finally, the performance of the proposed approach was evaluated on two datasets, achieving an average mIoU and accuracy of 82.7% and 91.3%. The proposed monitor was demonstrated to be effective, which holds significance in clinical applications.