Background

Present-day medical artificial intelligence (AI) models for medical visual question answering and report generation frequently function passively. Rather than simulating the dynamic nature of clinical workflow, they primarily respond to direct commands. This restricts their applicability in real-world healthcare settings, where complex reasoning and multi-turn dialogue are crucial.

Objective

This paper proposes a novel multimodal framework, based on Low-Rank Adaptation (LoRA), that simulates authentic doctor–patient interactions. The objective is to develop an AI assistant capable of participating in multi-turn diagnostic conversations with improved reasoning, interpreting radiological images, and addressing patient queries.

Methods

We used a curated and enriched version of the VQA-Med CLEF 2019 dataset to perform LoRA fine-tuning on a large-scale vision–language model. To train the model for both visual diagnosis and natural language interaction, the dataset was supplemented with simulated patient queries, radiological reports, and follow-up questions.

Results

Compared with baseline and prompt-based methods, our model demonstrated superior performance. It achieved higher accuracy and lower loss values while producing outputs that were more interpretable across both textual and visual domains. The model’s capacity to manage intricate, multi-turn diagnostic queries was further validated through structured assessments.

Conclusion

The proposed framework represents a significant advancement toward AI assistants designed for therapeutic settings. By integrating dialogue-based reasoning with multimodal understanding, it bridges the gap between passive AI tools and active diagnostic agents capable of interacting with patient data in real-world clinical scenarios.