A Preliminary Comparison of Efficient Net-B3 and ResNet-50 for Detection and Subtype Classification of Intracranial Hemorrhage on Head Computed Tomography

Abstract

Background & Objective: Intracranial hemorrhage (ICH) is one among leading neurological emergencies and diagnostic delay worsens outcomes. However, prompt radiological review is often unavailable in low-resource settings. Existing deep learning models for ICH detection largely prioritize accuracy over computational efficiency, restricting deployment on standard hardware. This is motivated by our own experience, working without institutional GPU access on free-tier cloud compute and encountered repeated training interruptions signalling infrastructure constraints faced by researchers in low-resource settings. Methodology: From 752,763 labeled axial slices, a random sample ofhemorrhage-positive and negatives were combined (n=212,933) and labeled for five subtypes: epidural (EDH), subdural (SDH), intra-parenchymal (IPH), intraventricular (IVH), subarachnoid (SAH), and an “any hemorrhage” category. Splitting was performed at the patient level with zero overlapping patients(n=148,753), validation (n=31,915) and test (n=32,265) sets. EfficientNet-B3 (10.7 million parameters) and ResNet-50 (25.6 million parameters), both were fine-tuned as multi-label classifiers using class-weighted binary cross-entropy loss. EfficientNet-B3 completed 1 epoch and ResNet-50 completed 2 due to recurring computational interruptions. Discrimination was assessed using AUC, sensitivity and specificity per subtype at a 0.5 threshold. We used publicly available, de-identified datasets from the Radiological Society of North America Intracranial Haemorrhage Detection (RSNA-IHD), so the Institutional Review Board (IRB) was exempted. Results: EfficientNet-B3 reached AUC 0.961 after 1 epoch and ResNet-50 reached AUC 0.959 after 2 epochs for “any hemorrhage”. Both discriminated IVH well (AUC 0.974-0.975) but performed weakest on epidural hemorrhage EDH (1.5% of training slices from the data); AUC 0.862-0.872, sensitivity 0.74. Given this training imbalance statistical comparison (DeLong's test) is planned after matched epoch training. Conclusion: Overall, this benchmark shows strong hemorrhage discrimination for ICH screening in resource-limited settings with a validated EDH sensitivity gap. Limitations include incomplete, unequal training models, balanced test-set prevalence (~50%) that will inflate precision relative to real prevalence (~5-15%); slice-leve evaluation and no external validation. EDH gap is filled by threshold recalibration, targeted oversampling and fine-tuning by increasedloss weightt during retraining. Key Words: Net-B3, RES-Net50, Intracranial Heamorhage, CT Brain.