Multiclass Intracranial Hemorrhage Detection and Confidence Aware Triage via Deep Ensemble Learning

Intracranial Hemorrhage is a critical medical condition characterized by bleeding within the skull, frequently resulting from traumatic brain injury, aneurysms, or other vascular abnormalities. A rapid and accurate diagnosis is essential, as delays can have life-threatening consequences. This paper presents a comprehensive automated deep learning framework that achieves robust multiclass classification of ICH from Computed Tomography scan images encompassing intraparenchymal, intraventricular, subarachnoid, subdural, and epidural hemorrhages, and integrates advanced techniques such as uncertainty estimation and ensemble learning to support clinical decision-making. Five state-of-the-art vision backbones - ConvNeXt, CoaTNet, Swin Transformer, ViT-B, and DeiT-S - were systematically benchmarked, and leveraging the three strongest networks we constructed a gradient-boosted meta-ensemble that achieved 0.948 accuracy, 0.970 micro-F1, and 0.967 macro-F1, raising macro recall to 0.957 and recovering the rare epidural subtype (F1 = 0.83). Beyond accuracy gains, the novelty of this work lies in a confidence-aware triage design: dropout was retained at inference to generate Monte Carlo uncertainty estimates, enabling calibrated probabilities (ECE = 2.8%) and a workflow in which high-confidence cases are automatically cleared while only ≈8% of uncertain slices are routed to radiologists. This selective strategy both reduces clinical workload and transparently communicates algorithmic doubt, a capability absent in prior ICH classifiers. The complete framework processes a head CT volume with uncertainty estimation in under four minutes on a single GPU, demonstrating deployment-ready throughput compatible with emergency radiology timelines. By combining systematic backbone benchmarking, meta-ensemble fusion, and slice-level uncertainty quantification, this study provides not only improved rare-subtype sensitivity but also a practical foundation for real-time, trustworthy Intracranial Hemorrhage triage in the acute-care setting.