Calibrating Label Distribution for Class-Imbalanced Barely-Supervised Knee Segmentation

Segmentation of 3D knee MR images is important for the assessment osteoarthritis. Like other medical data, volume-wise labeling expertise-demanded and time-consuming; hence semi-supervised learning (SSL), particularly barely-supervised learning, highly desirable training with insufficient labeled data. We observed that class imbalance problem severe in as cartilages only occupy 6% foreground volumes, situation becomes worse without sufficient To address above problem, we present a novel framework segmentation noisy imbalanced labels. Our leverages label distribution to encourage network put more effort into cartilage parts. Specifically, utilize 1) quantity modifying objective loss function class-aware weighted form 2) position constructing cropping probability mask crop sub-volumes areas from both unlabeled inputs. In addition, design dual uncertainty-aware sampling supervision enhance low-confident categories efficient unsupervised learning. Experiments show our proposed brings significant improvements by incorporating data alleviating imbalance. More importantly, method outperforms state-of-the-art SSL methods, demonstrating potential challenging setting. code available at https://github.com/xmed-lab/CLD-Semi .