Related Experiment Video
Updated: Sep 11, 2025
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
Automated Protocol Suggestions for Cranial MRI Examinations Using Locally Fine-tuned BERT Models
Christian Boschenriedter1, Christian Rubbert2, Marius Vach2
1Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital Düsseldorf, Heinrich-Heine-University Düsseldorf, Moorenstraße 5, 40225, Düsseldorf, Germany. christian.boschenriedter@med.uni-duesseldorf.de.
Bidirectional encoder representations from transformers (BERT)-based models can predict cranial magnetic resonance imaging (MRI) protocols from referral data. Fine-tuning with local language improved accuracy, showing potential for faster, standardized radiological protocol selection.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Radiology
Background:
- Selecting appropriate cranial MRI protocols is vital for accurate diagnosis and patient care.
- Current protocol selection relies on radiologist expertise, which can be time-consuming and variable.
- Inappropriate protocols risk diagnostic errors and prolonged scan times.
Purpose of the Study:
- To investigate the efficacy of BERT-based language models in suggesting cranial MRI protocols.
- To evaluate the performance of different BERT models, including medBERT.de, with local language fine-tuning.
- To assess the potential of AI to streamline and standardize radiological protocol selection.
Main Methods:
- Trained four BERT-based models (BERT, ModernBERT, GottBERT, medBERT.de) on 410 anonymized cranial MRI referrals.
- Classified referrals into nine protocol categories based on diagnoses, history, and clinical questions.
- Fine-tuned models using a local language dataset to enhance performance.
Main Results:
- The medBERT.de model, fine-tuned for local language, achieved the highest performance.
- This model correctly predicted 81% of protocols with a macro-F1 score of 0.71.
- Local language fine-tuning improved performance across all evaluated BERT models.
Conclusions:
- BERT-based language models show significant potential for predicting MRI protocols from referral data.
- AI-driven protocol selection can accelerate and standardize clinical workflows in radiology.
- This approach offers benefits even with limited training data, improving efficiency and accuracy.
