Uncertainty-Guided Semi-Supervised (UGSS) mean teacher framework for brain hemorrhage segmentation and volume quantification

Traumatic brain injury (TBI) is considered a critical neurological emergency with substantial morbidity and mortality rates across the world. Among significant neuropathological consequences of brain injuries, intracranial hemorrhage (ICH) stands out as a particularly urgent condition necessitating prompt diagnosis to avert life-threatening complications. However, the traditional manual approach to detecting and segmenting brain hemorrhages in CT scans is time-consuming and labor-intensive. This study proposes a fully automated uncertainty-guided framework for intracranial hemorrhage segmentation in brain CT scans. The framework is trained on a semi-supervised scheme that leverages both labeled and unlabeled data. Notably, when trained on 80% of labeled data, the semi-supervised framework yields an average Dice coefficient of 0.613 ± 0.01 and a Jaccard index of 0.441 ± 0.02. These metrics significantly exceed their supervised counterparts, which demonstrates the efficacy of the proposed methodology. Moreover, the proposed approach exhibits an overall accuracy of 89.03% in brain hemorrhage classification with a Cohen's Kappa value of 0.835, which indicates substantial agreement between the model's predictions and the ground truth labels. In addition to its capabilities in intracranial hemorrhage detection and localization, the proposed framework offers a robust estimation of hemorrhage volume and provides a comprehensive 3D volumetric view. The accuracy and reliability of the volume quantification approach are justified through a comprehensive qualitative and quantitative assessment, utilizing visualization techniques and a goodness-of-fit test (R²=0.837). In both instances, the method shows a notable alignment between the predicted hemorrhage volume and the actual hemorrhage volume. Thus, the proposed schemes of uncertainty-guided semi-supervised (UGSS) hemorrhage segmentation and volume quantification enhance model's applicability in clinical practice and research.