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

Automated Surgical Intervention Support Tool for Traumatic Brain Injury (ASIST-TBI)

https://doi.org/10.1148/ryai.230088

1.0

ot.htlaehytinu@witiW.rehpotsirhC

Supported by St Michael's Hospital Medical Services Association Innovation Fund. E.C. supported by the Odette Professorship in Artificial Intelligence for Medical Imaging, St Michael's Hospital, Unity Health Toronto.

Institutional research ethics board approval was obtained; informed consent was waived due to impracticality, minimal risk, and safeguards in place.

Vision Transformer-based model: CT scans split into fixed-size 1D patches with learnable spatial embeddings, passed through a transformer encoder (pretrained weights frozen), followed by a multilayer perceptron binary classification head.

Not publicly released; data provided by third parties (Ontario Trauma Registry, Unity Health Toronto).

Predicts need for acute neurosurgical intervention from initial trauma head CT to support triage and interfacility transfer prioritization, potentially improving resource allocation and patient care.

patients’ triage; clinical decision support systems.

Binary decision obtained by thresholding the classifier output; specific threshold not reported.

Assists clinician decision-making (decision support); does not automate clinical decisions.

Adults (≥18 years) presenting with acute traumatic brain injury undergoing noncontrast head CT in emergency/trauma settings to assess likelihood of neurosurgical intervention within 72 hours.

Preprocessed noncontrast head CT scans (NIfTI after DICOM conversion; registered to a standard head CT template).

Apply to acute, preoperative, noncontrast trauma head CT scans with adequate image quality; exclude postoperative, contrast-enhanced, duplicate navigation, and nondiagnostic motion scans; trained for adults ≥18 years; predictions pertain to interventions within 72 hours of scan acquisition.

Single-center development and testing with scans from a single CT manufacturer; potential bias from surgeon-level decision making in the ground truth; 72-hour intervention label may miss delayed interventions; misclassifications noted in cases with low GCS and small-volume hemorrhage (ICP monitoring), cases later exhibiting hemorrhage expansion, depressed/open skull fractures, and cases with large hemorrhages but DNR/poor neurologic status; fairness testing showed no significant differences by age or sex but lower GCS among misclassified cases; external validation pending.

Use as an adjunct for triage and decision support to prioritize transfers and neurosurgical consultation in acute TBI; clinical judgment and guidelines should prevail, especially in severe TBI with established care pathways.

Preprocessing and model architecture described; pretrained transformer encoder with frozen weights; additional training parameters provided in Appendix S1; code and model weights not provided.