NLM DIR Seminar Schedule

Abstract:

We introduce GPTRadScore, a groundbreaking evaluation framework for assessing multimodal large language models (LLMs) in CT imaging. Using GPT-4, GPTRadScore measures model performance in tasks like lesion localization, body part identification, and lesion typing. It outperforms traditional metrics such as BLEU and ROUGE, aligning closely with expert clinician assessments. Fine-tuning with specialized datasets significantly boosts performance, as demonstrated by RadFM’s notable improvements in accuracy.
To support the development of AI in CT imaging, we also present CT-Bench, a comprehensive dataset containing 20,335 annotated lesions from 7,795 patient studies. Accompanied by high-quality, GPT-4-enhanced textual descriptions and a visual question-answering (VQA) benchmark with 2,850 QA pairs, CT-Bench enables targeted training and evaluation of AI models for lesion description, localization, and diagnostic reasoning.
Together, GPTRadScore and CT-Bench provide powerful tools to advance multimodal AI, setting new standards for evaluation, training, and performance in CT imaging analysis.