https://atlas.rsna.org/schemas/2025-11/model.json

Examination-Level Supervision for Deep Learning–based Intracranial Hemorrhage Detection on Head CT Scans

https://pubmed.ncbi.nlm.nih.gov/38294324/

1.0

Corresponding author: Jacopo Teneggi (email: ude.uhj@1ggenetj)

National Science Foundation CAREER Award Computing and Communication Foundations 2239787.

Retrospective study using public data; acknowledged as nonhuman subjects research by the University of Maryland Baltimore institutional review board.

Attention-based Deep Multiple Instance Learning model with a shared convolutional neural network feature extractor (ResNet-18) for 2D slices and a fully connected sigmoid classifier; strong learner operates at image level; weak learner aggregates bag-level (exam-level) predictions via attention.

Authors state: All code will be released. (No URL provided in article)

Automated detection and triage of intracranial hemorrhage on head CT; maintains or improves performance using examination-level labels, potentially reducing radiologist annotation burden by ~35x.

Clinical decision support and triage (flagging exams/images with suspected ICH); workflow optimization by prioritizing studies and localizing suspicious sections.

Not explicitly specified; AUC used for evaluation. Section-level selection for WL based on attention weights or Shapley values thresholding.

Decision support; provides automated exam-level classification and suggested section/pixel-level localizations via saliency maps to support radiologist review.

Detection of intracranial hemorrhage in patients undergoing unenhanced head CT examinations in acute care settings; intended for analysis of CT volumes of arbitrary number of sections.

Unenhanced head CT examinations treated as bags of 2D images (axial sections); images windowed with brain settings (WW=80, WL=40) and min–max normalized.

Train with focal loss to address class imbalance; use exam-level labels for weak learner or image-level labels for strong learner; for WL section-level localization, select images by attention weights or Shapley values (h-Shap). Data split 80/20 at exam level on RSNA 2019 dataset; externally test on CQ500 and CT-ICH.

Study limited to ICH detection on head CT; some image-level annotations still required for section/pixel-level evaluation; no pixel-level training (no segmentation labels in RSNA); generalization assessed on two external datasets only; other architectures (e.g., vision transformers) not evaluated.

For large datasets (m ≥ ~5,000–10,000 labeled exams), use examination-level weak supervision (MIL) to achieve strong exam-level and section-level performance with substantially fewer labels; accompany outputs with saliency maps to aid review.

Training/validation on public RSNA 2019 dataset with random 80/20 split at exam level; external validation on public CQ500 and CT-ICH datasets; preprocessing (brain windowing and normalization) described; hyperparameters and augmentation in Appendix S1; code to be released.