Role of Large Language Models for Suggesting Nerve Involvement in Upper Limbs MRI Reports with Muscle Denervation...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Neuromuscular disease diagnosis

Background:

• Identifying specific peripheral nerves (PNs) involved in upper limb muscle denervation from MRI reports is complex.
• Large language models (LLMs) offer potential for automating this diagnostic process.

Purpose of the Study:

• To develop, compare, and validate LLMs for automatically identifying relationships between denervated muscles and PNs in upper limb MRI reports.
• To assess the performance of various LLMs, including BERT, DistilBERT, mBART, RoBERTa, and Medical-ELECTRA, in this task.

Main Methods:

• Retrospective analysis of 300 Spanish upper limb MRI reports with muscle denervation signs (2018-2024).
• Manual annotation of affected PNs (median, ulnar, radial, axillary, suprascapular) by an expert radiologist.
• Fine-tuning and evaluation of multiple LLMs, coupled with an automatic majority voting system for consolidated predictions.

Main Results:

• The voting system achieved high F1 scores: 0.88 (median), 1.00 (ulnar), and 0.90 (radial).
• Medical-ELECTRA demonstrated strong performance (F1 > 0.82) for axillary and suprascapular nerves.
• mBART showed lower performance, with an F1 score of 0.38 for the median nerve.

Conclusions:

• The developed voting system generally surpasses individual LLMs in pinpointing PNs associated with muscle denervation on upper limb MRIs.
• This AI-driven approach can aid radiologists by suggesting implicated PNs in their reports, improving diagnostic efficiency.