Self-Training Segmentation of COVID-19 Lesions

With the recent advancements in computer technology and its integration into healthcare, deep learning-based techniques have significantly improved the diagnosis and treatment of various diseases. The success of these approaches depends on large-scale annotated datasets. However, collecting and annotating large amounts of medical data is challenging due to concerns such as patient privacy, data anonymization, and the requirement of domain expertise during the annotation process. This study aims to improve segmentation performance on medical datasets with limited labeled data using a semi-supervised learning approach. Semi-supervised training generates predictions for unlabeled data and incorporates this information into the model to enhance its performance. In the study, computed tomography (CT) images of COVID-19 patients were used, and a self-training algorithm was used to improve lesion segmentation. The Monte Carlo Dropout method was applied to increase the reliability of pseudo-labels, which significantly increased the model’s performance. The results demonstrate the effectiveness of self-training-based approaches in medical imaging scenarios with limited annotated data.