Medical Knowledge-Guided Semi-Supervised Bi-Ventricular Segmentation

Pseudo-labeling is a well-studied semi-supervised learning approach that generates artificial labels for unlabeled data based on the predictions of an initial model trained on labeled data. Although pseudo-labeling is an effective approach for a wide range of tasks, incorporating physical knowledge concentrated on a particular organ (such as cardiac structures), can outperform the general strategy employed in pseudo-labeling. In this work, we propose to integrate different physical (medical) properties into the semi-supervised bi-ventricular segmentation task. We incorporate this knowledge as regularization terms in the loss function, uncertainty criteria for assessing the predictions, and pseudo-label modification methods. Our extracted properties are based on the physical characteristics of the ventricles and are robust to modality changes. We validated our method using the ACDC and SCD datasets. Numerical measurements confirm the success of the proposed approach. The implementation of our work is available at “https://github.com/behnamrahmati/MedicalGuided”