Towards Explainability in Deep Learning for Detection of Five Major Intracranial Hemorrhage Subtypes on Head CT Using Multi-window DICOM Imaging and Patient-Level Cross-Validation.
An AI system detects brain bleeding on CT scans with 97% accuracy, though rare subtypes remain challenging and real-world validation is needed.
An explainable ResNet34 deep learning framework trained on DICOM-native three-window CT (brain/subdural/bone) achieved robust multi-label ICH detection (ROC AUC 0.97+) with patient-level cross-validation on 18,938 patients from the RSNA 2019 challenge, accompanied by multi-method attribution analysis supporting model interpretability. Performance on rare subtypes (particularly epidural hemorrhage) remains limited, and prospective external validation is required.
What the study was
- Study design
- Retrospective deep learning validation (RSNA 2019 ICH challenge dataset; 3-fold patient-level cross-validation)
- Population
- CT scans from 18,938 patients (RSNA 2019 ICH detection challenge)
- Sample size
- 18938
- Category
- Diagnostics
- Maturity
- Exploratory
- Journal
- Journal of Imaging Informatics in Medicine
Why it surfaced
Large public-dataset DICOM-native explainable AI for ICH detection; strong ROC AUC with multi-method explainability; pending prospective validation. J Imaging Inform Med.
A plain-language summary of published research — not medical advice. Talk to a clinician about your own care.