Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Non-contrast Head CT Foundation Model for Comprehensive Neuro-Trauma Triage

Youngjin Yoo, Bogdan Georgescu, Yanbo Zhang, Sasa Grbic, Han Liu, Gabriela D. Aldea, Thomas J. Re, Jyotipriya Das, Poikavila Ullaskrishnan, Eva Eibenberger, Andrei Chekkoury, Uttam K. Bodanapally, Savvas Nicolaou, Pina C. Sanelli, Thomas J. Schroeppel, Yvonne W. Lui, Eli Gibson

TL;DR

This work addresses the urgent need for rapid, accurate neuro-trauma triage in emergency head CT by developing a 3D foundation model (CNTD-Net) that combines LLM-generated multi-label annotations with task-specific pretraining on hemorrhage segmentation and brain anatomy parcellation. The model is enhanced through multimodal finetuning to form DeepCNTD-Net, integrating hemorrhage and anatomy features for comprehensive detection across 16 neuro-trauma findings, achieving an average AUC of $0.861$ and outperforming CT-CLIP on key metrics. LLM labeling demonstrates high accuracy on major findings and reasonable performance on rarer conditions, enabling scalable annotation without extensive manual labeling. The results show strong generalization to external datasets (e.g., CQ500) and highlight the value of domain-specific priors in medical-imaging foundation models, paving the way for clinical integration to alleviate radiologist workload and improve triage efficiency.

Abstract

Recent advancements in AI and medical imaging offer transformative potential in emergency head CT interpretation for reducing assessment times and improving accuracy in the face of an increasing request of such scans and a global shortage in radiologists. This study introduces a 3D foundation model for detecting diverse neuro-trauma findings with high accuracy and efficiency. Using large language models (LLMs) for automatic labeling, we generated comprehensive multi-label annotations for critical conditions. Our approach involved pretraining neural networks for hemorrhage subtype segmentation and brain anatomy parcellation, which were integrated into a pretrained comprehensive neuro-trauma detection network through multimodal fine-tuning. Performance evaluation against expert annotations and comparison with CT-CLIP demonstrated strong triage accuracy across major neuro-trauma findings, such as hemorrhage and midline shift, as well as less frequent critical conditions such as cerebral edema and arterial hyperdensity. The integration of neuro-specific features significantly enhanced diagnostic capabilities, achieving an average AUC of 0.861 for 16 neuro-trauma conditions. This work advances foundation models in medical imaging, serving as a benchmark for future AI-assisted neuro-trauma diagnostics in emergency radiology.

A Non-contrast Head CT Foundation Model for Comprehensive Neuro-Trauma Triage

TL;DR

This work addresses the urgent need for rapid, accurate neuro-trauma triage in emergency head CT by developing a 3D foundation model (CNTD-Net) that combines LLM-generated multi-label annotations with task-specific pretraining on hemorrhage segmentation and brain anatomy parcellation. The model is enhanced through multimodal finetuning to form DeepCNTD-Net, integrating hemorrhage and anatomy features for comprehensive detection across 16 neuro-trauma findings, achieving an average AUC of and outperforming CT-CLIP on key metrics. LLM labeling demonstrates high accuracy on major findings and reasonable performance on rarer conditions, enabling scalable annotation without extensive manual labeling. The results show strong generalization to external datasets (e.g., CQ500) and highlight the value of domain-specific priors in medical-imaging foundation models, paving the way for clinical integration to alleviate radiologist workload and improve triage efficiency.

Abstract

Recent advancements in AI and medical imaging offer transformative potential in emergency head CT interpretation for reducing assessment times and improving accuracy in the face of an increasing request of such scans and a global shortage in radiologists. This study introduces a 3D foundation model for detecting diverse neuro-trauma findings with high accuracy and efficiency. Using large language models (LLMs) for automatic labeling, we generated comprehensive multi-label annotations for critical conditions. Our approach involved pretraining neural networks for hemorrhage subtype segmentation and brain anatomy parcellation, which were integrated into a pretrained comprehensive neuro-trauma detection network through multimodal fine-tuning. Performance evaluation against expert annotations and comparison with CT-CLIP demonstrated strong triage accuracy across major neuro-trauma findings, such as hemorrhage and midline shift, as well as less frequent critical conditions such as cerebral edema and arterial hyperdensity. The integration of neuro-specific features significantly enhanced diagnostic capabilities, achieving an average AUC of 0.861 for 16 neuro-trauma conditions. This work advances foundation models in medical imaging, serving as a benchmark for future AI-assisted neuro-trauma diagnostics in emergency radiology.

Paper Structure

This paper contains 11 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. The comprehensive neuro-trauma detection foundation model
  3. Automatic comprehensive labeling
  4. Task-specific pretraining
  5. Building a foundation model with multimodal finetuning
  6. Results and Discussion
  7. Dataset and Preprocessing
  8. LLM labeling accuracy against manual expert label
  9. Comprehensive neuro-trauma detection performance
  10. Conclusion
  11. Disclaimer

Figures (4)

  • Figure 1: Overview of the head CT foundation model for neuro-trauma triage.
  • Figure 2: A comprehensive list of neuro-trauma findings that encompass the critical conditions requiring immediate clinical attention in trauma emergency centers curated by a neuro-radiologist and their prevalence in our dataset.
  • Figure 3: Individual detection performance (AUC) for six major critical findings and all 16 neuro-trauma conditions. CNTD-Net was evaluated exclusively for the six major findings.
  • Figure 4: Examples of comprehensive neuro-trauma triage. "POS" indicates a positive case, while "NEG" denotes a negative case, both based on the LLM-gerated multi-labels, and AI triage score is between 0.0 and 1.0.