ABSTRACT NUMBER: ESOC2026A1207 AUTOMATED CLASSIFICATION OF INTRACRANIAL HEMORRHAGE SUBTYPE FROM HEAD CT SCANS: A COMPARISON OF TWO DEEP LEARNING APPROACHES

Abstract Background and aims Spontaneous and trauma-related intracranial hemorrhage (ICH) are acute brain injuries that require urgent care. Non-contrast computed tomography (NCCT) is the primary type of imaging to support diagnosis and identify the ICH subtype. Due to the time-sensitivity, automated characterization of ICH subtypes may offer clinical insight and acute diagnostic support. Existing ICH deep learning classifiers focus on binary classification, while our method focuses on multi-label classification. Methods RSNA ICH Detection Challenge 2019 CT imaging data (n=22061) were reconstructed and split into n=17649 training, n=2206 validation, and n=2206 test sets. Ground truth ICH subtypes included: epidural (EDH), acute subdural (ASDH), subarachnoid (SAH), intraparenchymal (IPH) and intraventricular (IVH) hemorrhage. Figure 1 illustrates the ICH subtypes (arrows). Two deep learning approaches were trained and tested: 1) 3D-DenseNet, a parameter-efficient variant of a Convolutional Neural Network, 2) DINOv3GRU, which uses a pretrained vision transformer combined with a Gated Recurrent Unit (GRU). Results Figure 2 shows the area-under-the-curve (AUCs) classification results over 5 folds of the data. Validation and test results are provided for the mean across all subtypes, and individual ICH subtypes. Mann-Whitney U tests (DenseNet vs DINOv3GRU) revealed a significant AUC difference for the validation (W=0, p=0.008) and test set (W=0, p=0.008). Conclusions Two image-based ICH classification approaches were developed with performance that exceeded an AUC of 0.90 for every bleed subtype. Mann-Whitney U tests illustrated significantly higher AUCs for DINOv3GRU. These findings demonstrate the potential of deep learning models for ICH classification, in particular for time-sensitive clinical settings and further management. Conflict of interest All authors have nothing to disclose Figure 1 - belongs to Methods Figure 2 - belongs to Results