A hybrid approach for enhancing pseudo-labeling in medical images through pseudo-label refinement

Segmentation of medical images is critical for the evaluation, diagnosis, and treatment of various medical conditions. While deep learning-based approaches are the dominant methodology, they rely heavily on abundant labeled data and face significant challenges when data is limited. Semi-supervised learning methods mitigate this issue but there are still some challenges associated with them. Additionally, these approaches can be improved specifically for medical images considering their unique properties (e.g., smooth boundaries). In this work, we adapt and enhance the well-established pseudo-labeling approach specifically for medical image segmentation. Our exploration consists of modifying the network's loss function, pruning the pseudo-labels, and refining pseudo-labels by integrating traditional image processing methods with semi-supervised learning. This integration enables traditional segmentation techniques to complement deep semi-supervised methods, particularly in capturing fine edges where deep models often struggle. It also incorporates the smoothness of the edges in the segmentation and achieves a balance between deep learning and traditional methods through tunable parameters. Moreover, to address the problem of noisy or unreliable pseudo-labels, we utilize uncertainty-based pixel-level and image-level pruning of the pseudo-labels using a specific loss function, thereby improving the accuracy and robustness of the segmentation. We evaluated our approach on three different datasets from two imaging modalities (CT and MRI) and demonstrated its superior performance, highlighting its accuracy and robustness in the presence of limited labeled data. With only 15% of the labeled data, on the Sunnybrook Cardiac dataset, our approaches increased endocardium segmentation accuracy from 82.1% to 87.5%, and epicardium segmentation from 82.5% to 86.7%. On the COVID-19 CT lung and infection segmentation dataset, our approach improved left lung segmentation accuracy from 72.5% to 79.3%, and right lung segmentation from 75.8% to 81.6% when using only 15% of labeled data. On the Automated Cardiac Diagnostic Challenge dataset, with just 10% of labeled data, our approach increased endocardium segmentation from 91% to 93.7%, myocardium from 69.8% to 74.5%, and right ventricle from 76.7% to 82.1%. Our codes will be published in https://github.com/behnam-rahmati .