Large Language Model Based Identification of Brain MRI Sequences

Motivation: Classifying MRI sequences automatically is critical for developing labeled datasets for deep learning applications in medical imaging. Goal(s): The study evaluates large language models(LLMs) performance in classifying MRI sequences, comparing them to current methods like CNNs and string-matching, focusing on accuracy and interpretability. Approach: We applied a GPT-4-based LLM to classify 1490 brain MRI sequences from UCSF and compared to CNN and string-matching classifiers using sensitivity, specificity, and accuracy. Results: The LLM classifier outperformed both CNN and string-matching methods, achieving 0.83 accuracy, with high sensitivity and specificity across sequence types. Its interpretability offered additional insights, improving classification transparency and minimizing false positives. Impact: LLMs provide a more accurate and interpretable approach for MRI sequence classification, offering clinicians and researchers a more reliable tool. This could enhance research workflows, reduce manual labeling time, and allow for more robust deep learning models in medical imaging.