Synthetic data trained open-source language models are feasible alternatives to proprietary models for radiology reporting

Abstract

The study assessed the feasibility of using synthetic data to fine-tune various open-source LLMs for free text to structured data conversation in radiology, comparing their performance with GPT models. A training set of 3000 synthetic thyroid nodule dictations was generated to train six open-source models (Starcoderbase-1B, Starcoderbase-3B, Mistral-7B, Llama-3-8B, Llama-2-13B, and Yi-34B). ACR TI-RADS template was the target model output. The model performance was tested on 50 thyroid nodule dictations from MIMIC-III patient dataset and compared against 0-shot, 1-shot, and 5-shot performance of GPT-3.5 and GPT-4. GPT-4 5-shot and Yi-34B showed the highest performance with no statistically significant difference between the models. Various open models outperformed GPT models with statistical significance. Overall, models trained with synthetic data showed performance comparable to GPT models in structured text conversion in our study. Given privacy preserving advantages, open LLMs can be utilized as a viable alternative to proprietary GPT models.