A randomized multi-window 3D deep learning approach for intracranial hemorrhage detection on non-contrast head CT

Intracranial hemorrhage (ICH) is an emergency clinical condition that requires rapid diagnosis. In patients with suspected ICH, head computed tomography (CT) is the first-line imaging modality because it provides fast results. In recent years, interest has increased in 3D deep learning approaches for automated scan-level ICH detection. However, real-world deployability is limited by high computational cost, memory requirements, and inference latency. In this study, we present Lite3DNet, a lightweight architecture that uses randomized multi-window 3D input. We used the Radiological Society of North America (RSNA) ICH Detection dataset, the largest public dataset related to ICH. DICOM series were converted into 3D volumes, data cleaning steps were applied, and single-window, three-window, and training-time window jittering strategies were investigated. The proposed architecture was trained using 5-fold cross-validation and evaluated on a held-out internal test set and two external test sets, CQ500 and MosMedData CT Hemorrhage (MosMed). The results show that Lite3DNet achieves high discriminative performance. In ensemble evaluation, the model reached an AUC of 0.986 on the RSNA ICH Detection held-out internal test set, 0.956 on the external CQ500 test set and 0.967 on the external MosMed test set. With 1.189 million parameters and 19.84 ms GPU inference latency, the model delivers low computational cost while remaining suitable for clinical triage applications that require rapid inference.