An Explainable Multi-Model Ensemble for Brain Hemorrhage Detection and Subtype Classification Using 34,366 NCCT Images

Background: Brain hemorrhage is a critical neurological emergency and it is essential to quickly and accurately identify the type of bleeding if treatment is to be properly guided and patient outcomes are to be improved. Even though non-contrast computed tomography (NCCT) is still the main imaging technique used for the initial evaluation, diagnostic difficulties continue due to overlapping radiological patterns and the limitations involved in manual interpretation. The purpose of this study was to create and validate an ensemble deep learning approach for the automated multi-class classification of brain hemorrhage subtypes based on NCCT images. Methods: Transfer learning was applied using three convolutional neural network architectures such as InceptionV3, Xception, and InceptionResNetV2. The models were trained independently and combined using an unweighted soft-voting strategy. The framework was evaluated on 34,366 NCCT images belonging to six diagnostic categories: EDH, SDH, SAH, IPH, IVH, and normal findings. Results: Xception achieved the highest performance among the individual models, whereas the ensemble produced the best overall results. Using the final test-set results, the ensemble achieved an accuracy of 95.35%, with macro-averaged precision, recall, and F1-score of approximately 95%. Grad-CAM visualizations showed that the model frequently attended to anatomically plausible regions associated with the predicted hemorrhage categories. Conclusions: The findings suggest that combining complementary CNN architectures may improve internal test-set performance for six-class NCCT image classification. However, because the study used an image-level split from a single public dataset without external validation, the proposed framework should be considered a research-stage decision-support approach rather than a clinically validated diagnostic system.