Development of a new method M2LC for few-slice-annotated MRI image segmentation and validation of its performance on multiple datasets

We designed a semi-supervised method that effectively utilized few-slice-annotated data for efficient MRI segmentation. We investigate a weakly-supervised scenario for medical image segmentation named as few-slice-annotated segmentation, where only a limited number of slices within each 3D volume are annotated for model training. To address the significant label scarcity in this scenario, we present a method called Maximum Minimum Entropy Cross Supervision and Label Correction (M2LC). M2LC initially trains semi-supervised 2D segmentation networks with a Siamese structure by transforming 3D volumes into 2D slices. During training, we apply entropy minimum and maximum constraints to the two branches of the network, respectively, and employ mutual supervision. Subsequently, we use the minimum entropy network to generate 2D pseudo-labels, which are then assembled back into 3D volumes and refined via a 3D label correction process. We validated the M2LC method on several datasets, including BraTs, ADCD, and a bladder cancer dataset collected from the Shanghai General Hospital. Compared with fully-supervised methods, M2LC achieves 83 % labeling accuracy with 0.6 % annotated slices on the BraTs dataset, and 79 % labeling accuracy with 1.2 % annotated slices on the ADCD dataset. On the bladder cancer dataset from Shanghai General Hospital, M2LC achieves labeling accuracies of 95.79 % and 99.13 % for the bladder wall and bladder tumor regions, respectively, with 23.17 % and 24.51 % annotated slices. This study successfully developed a semi-supervised deep learning method for MRI image segmentation under partial slice annotation conditions, achieving state-of-the-art performance across multiple datasets.