A hybrid deep learning framework for multi-class ICH detection and severity assessment from CT scans

Abstract

Intracranial hemorrhage (ICH) is a critical neurological condition that requires rapid and accurate assessment for effective clinical management. Analysis of head CT scans using automated systems can be used to help clinicians as it enhances the reliability of detections and can be used to prioritize in the emergency room. Nevertheless, the current approaches tend to detect hemorrhage only and use considerable annotated datasets, whereas the severity assessment is still underresearched because of the lack of ground-truth labels. In this thesis, we will present a hybrid deep learning model to detect multi-class intracranial hemorrhage and severity stratification by proxy as an element of clinical metadata and head CT images. The framework utilizes an EfficientNet-B0 backbone, to obtain slice level visual features, which are pooled in order to obtain low contextual information between slices. The feature fusion is used to add more contextual cues by incorporating patient level metadata. The model is trained under a multi-task learning environment, where it concurrently detects the subtypes of hemorrhage and classifies the category of the severity. Since the CQ500 dataset lacks explicit severity annotations, proxy hemorrhage volume estimates and dataset specific quantile thresholds are used to create severity labels in a weakly supervised fashion. This allows relative stratification of severity in the dataset as opposed to clinical grading of severity. Transfer learning is used to pretrain on the RSNA Intracranial Hemorrhage dataset and then fine-tune on CQ500 to enhance generalization with limited data. Experimental results demonstrate competitive patient-level detection performance across multiple hemorrhage subtypes, achieving high ROC-AUC values on CQ500. The severity classification task achieves moderate and consistent performance, with most errors occurring between adjacent severity levels, reflecting the continuous nature of hemorrhage extent and the proxy-based labeling scheme. Overall, this work demonstrates the feasibility of combining multimodal data, weak supervision, and multi-task learning for intracranial hemorrhage analysis under constrained annotation settings. While not intended for direct clinical deployment, the proposed framework provides a foundation for future research incorporating expert-annotated severity labels and external validation.