Fig. 5

Architecture and performance of PathChat, a multimodal AI pathology assistant. Detailed representation of PathChat's architecture, combining a UNI visual-language pretrained model with Llama 2 LLM, fine-tuned on 456,916 instructions. The figure highlights PathChat's superior performance in multiple-choice diagnostic questions (78.1–89.5% accuracy) and open-ended question answering (78.7% accuracy), significantly outperforming other models like LLaVA 1.5, LLaVA-Med, and ChatGPT-4