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

Semi-supervised Learning for Generalizable Intracranial Hemorrhage Detection and Segmentation

https://doi.org/10.1148/ryai.230077

1.0

Esther L. Yuh (ude.fscu@huy.rehtse)

Supported by the California Institute to Advance Precision Medicine (CIAPM).

Retrospective study approved by the UCSF IRB; HIPAA compliant; consent waived per 45 CFR 46.116(d) and FDA guidance.

PatchFCN: fully convolutional neural network with a Dilated ResNet-38 backbone; two-branch design for classification and segmentation; implemented in PyTorch using DRN codebase.

Research model; implemented in PyTorch using public DRN backbone code (https://github.com/fyu/drn). PatchFCN details: https://doi.org/10.1073/pnas.1908021116.

Automated detection and pixel-level segmentation of intracranial hemorrhage on head CT to support triage, localization, and quantification for downstream clinical management.

Patients’ triage; clinical decision support systems.

Semi-supervised ranker set top 10% of unlabeled images (C=10) as positive. Pixel pseudo-labeling thresholds: probability >0.7 positive, <0.3 negative, in-between ignored. Example operating point: sensitivity 0.920 with specificity 0.804 (semi-supervised) on CQ500.

Automates exam-level classification (hemorrhage vs no hemorrhage) and pixel-wise segmentation; intended to support, not replace, clinical decision-making.

Detection and segmentation of intracranial hemorrhage on noncontrast head CT examinations; intended for use on head CT images in radiology workflows to aid assessment of acute traumatic brain injury.

Noncontrast head CT images (DICOM); model operates on image patches during training and inference.

Run model on head CT; outputs exam-level hemorrhage probability and a pixel-wise hemorrhage probability map/mask. For training, semi-supervised noisy-student paradigm with data augmentation (contrast transformations; head size/aspect ratio) and a ranker for pseudo-label selection.

Higher computational requirements due to large unlabeled corpus; retraining needed as datasets grow. Labeled training data from a single institution and single vendor; training set hemorrhage prevalence higher than real world. Pseudo labels may propagate false positives without ranker. Generalization evaluated primarily on CQ500 (India).

Use semi-supervised noisy-student training with ranker and data augmentation to improve generalization for intracranial hemorrhage detection/segmentation on out-of-distribution head CT datasets.

Model details provided: training for 600 epochs, step number 240, batch size 16, crop size 240; minibatch mixing ratio Atlantis:Kaggle-25K = 0.6:0.4; single NVIDIA Tesla V100 GPU; performance estimated with 1000x bootstrapping; test-retest segmentation repeated three times showing consistent superiority.

Single NVIDIA Tesla V100 GPU used; training 600 epochs; runtime/energy not reported.