Semi-supervised meniscus segmentation with adaptive weighting strategies in a limited labeled data environment

Meniscus allograft transplantation is a primary method for treating damaged meniscus, and accurate segmentation is essential to determine the appropriate size and shape of the graft. However, meniscus segmentation is challenging due to the similar intensity between the meniscus and surrounding structures, as well as the variability in its shape and thickness. In this paper, we aim to enhance the generalization performance of models on knee MR images and improve segmentation accuracy by effectively accounting for the diverse shapes and thicknesses of the meniscus. To this end, we propose a semi-supervised learning model capable of effective training even with limited labeled data. Experimental results demonstrate that the proposed method achieved the highest performance with a recall of 93.28%, an F1-score of 90.02%, and a balanced accuracy of 96.42%, and by addressing the under-segmentation issue. it achieved a recall that is 0.8%p higher than that of the supervised learning model trained with 83 labeled data.