A multi-task deep learning pipeline for classification, detection, and weakly supervised 3D segmentation of intraparenchymal hematoma on brain CT

This study developed a multi-task deep learning pipeline for the automated assessment of acute intracranial hemorrhage and perihematomal edema on non-contrast brain computed tomography. Acute hemorrhage remains a major neurological emergency, and rapid and reliable image interpretation is essential for timely management. To address this clinical need, the proposed framework combined three complementary tasks in a single workflow: classification to identify the presence and subtype of hemorrhage, detection of perihematomal edema, and three-dimensional segmentation of intraparenchymal hematoma (IPH) using a weakly supervised strategy with pseudo-labels derived from edema masks. A total of 10,922 images were analyzed: 6,000 RSNA images, 728 PHE-SICH-CT-IDS slices from 120 patients, and 4,194 BHSD images. These models achieved high sensitivity (0.9584 ± 0.0068) for hemorrhage detection, robust performance (AUC, 0.9330 ± 0.0072) in differentiating subtypes, and reliable identification of edema (detection rate, 0.9873). Segmentation accuracy was excellent for IPH (Dice Similarity Coefficient [DSC], 0.9803 ± 0.0069) and moderate for edema (0.4569 ± 0.1809), with external validation confirming generalizability across centers (0.7549 ± 0.1143) for IPH. By integrating classification, detection, and segmentation, this pipeline demonstrates the potential of deep learning to provide accurate and scalable support for clinical decision-making, enhance diagnostic confidence, and streamline clinical workflow in the acute care setting.