The model was trained in Google Colab on a free Tesla T4 using 82 patients, 2,501 slices, and 318 hemorrhage-positive slices from an open Kaggle dataset. It used EfficientNet-B0 with 320x320 inputs and five patient-grouped validation folds.
The full input, combining brain and bone windows with a contrast-stretched brain channel, reached slice-level AUC 0.931 and patient-level AUC 0.914. An initial fold had reported AUC 0.963, but the full cross-validation result was lower. A bone-window-only model reached AUC 0.861; the author tested fractures and slice position as possible confounders but did not identify the source of that signal.
Adding a segmentation decoder left AUC nearly unchanged, at 0.932, while average precision rose from 0.699 to 0.795. At a target 95% slice-level sensitivity, specificity was 0.70 to 0.75; patient-level targets were not sustained, with two to three missed patients among 36 positive cases. The source does not establish clinical validity, diagnostic suitability, or regulatory clearance.