SigPhi-Med: A lightweight vision-language assistant for biomedicine.

Abstract

Recent advancements in general multimodal large language models (MLLMs) have led to substantial improvements in the performance of biomedical MLLMs across diverse medical tasks, exhibiting significant transformative potential. However, the large number of parameters in MLLMs necessitates substantial computational resources during both training and inference stages, thereby limiting their feasibility in resource-constrained clinical settings. This study aims to develop a lightweight biomedical multimodal small language model (MSLM) to mitigate this limitation. We replaced the large language model (LLM) in MLLMs with the small language model (SLM), resulting in a significant reduction in the number of parameters. To ensure that the model maintains strong performance on biomedical tasks, we systematically analyzed the effects of key components of biomedical MSLMs, including the SLM, vision encoder, training strategy, and training data, on model performance. Based on these analyses, we implemented specific optimizations for the model. Experiments demonstrate that the performance of biomedical MSLMs is significantly influenced by the parameter count of the SLM component, the pre-training strategy and resolution of the vision encoder component, and both the quality and quantity of the training data. Compared to several state-of-the-art models, including LLaVA-Med-v1.5 (7B), LLaVA-Med (13B) and Med-MoE (2.7B × 4), our optimized model, SigPhi-Med, with only 4.2B parameters, achieves significantly superior overall performance across the VQA-RAD, SLAKE, and Path-VQA medical visual question-answering (VQA) benchmarks. This study highlights the significant potential of biomedical MSLMs in biomedical applications, presenting a more cost-effective approach for deploying AI assistants in healthcare settings. Additionally, our analysis of MSLMs key components provides valuable insights for their development in other specialized domains. Our code is available at https://github.com/NyKxo1/SigPhi-Med.