MTW-ICHNet: Multi-task Weakly Supervised Learning with Enhanced Feature Descriptor Learning for Intracranial Hemorrhage Diagnosis.

Current research on intracerebral hemorrhage (ICH) detection faces challenges regarding feature utilization efficiency and multi-task collaboration under weakly supervised learning (WSL). While WSL reduces dependence on extensive labeled medical data, existing methods exhibit limitations in leveraging discriminative features and coordinating complementary tasks. To address these limitations, this study introduces MTW-ICHNet, a multi-task WSL network model that integrates a feature descriptor with collaborative task optimization. Feature enhancement techniques are implemented to refine extracted features during WSL training, thereby optimizing the use of limited annotations and improving discriminative capability. Simultaneously, interdependent sub-tasks-including lesion localization and category recognition-are jointly optimized through a unified architecture that facilitates cross-task knowledge sharing. The framework achieves accuracies of 98.7% and 97.5% for hemorrhage classification and lesion localization tasks, respectively. Experimental results demonstrate the enhanced performance in ICH image recognition through effective feature refinement and task collaboration. The proposed method shows potential for clinical applications by providing accurate diagnostic references to guide patient-specific treatment strategies, particularly in resource-constrained environments where comprehensive medical annotations are scarce.