The researcher developed an AI model to detect intracranial hemorrhages using a dataset of 82 patients, including 36 cases with hemorrhages and 46 without. The model was trained on 2,500 CT slices, 318 of which contained hemorrhages, using a five-fold cross-validation approach.
The baseline model achieved an AUC of 0.931. Incorporating segmentation improved the average precision from 0.70 to 0.83, though it did not significantly change the AUC. The model utilized EfficientNet-B0 and was trained on images converted to JPG format, which limited the availability of Hounsfield units.
Performance evaluation revealed that the model struggled to maintain a 95% recall rate at the patient level. When tested, the system missed between 2 and 3 patients out of 36 and produced between 8 and 18 false positives per 46 negative studies.
The study highlights that bone window analysis contributed to detection, achieving an AUC of 0.86, though the exact signal source remains unclear. The researcher noted that the project is an educational prototype and lacks clinical validation, emphasizing that results cannot be generalized to real-world medical diagnostics.
