[2507.17303] A Versatile Pathology Co-pilot via Reasoning Enhanced Multimodal Large Language Model

View the PDF file for the paper entitled “Good Multiple Paths Common User through a large multimedia language model, by ZHE XU and 15 other authors

PDF HTML (experimental) view

a summary:Melms appeared as strong Melms as strong tools for mathematical pathology, providing unprecedented opportunities to integrate pathological images with the context of the language to analyze a comprehensive diagnosis. These models have a special promise to automate complex tasks that traditionally require an expert interpretation of pathologists. However, the current MLLM methods in pathology show greatly restricted thinking capabilities, due primarily to their dependence on expensive explanatory comments. In addition, the current methods remain limited to the Simplex application to answer visual questions (VQA) at the ROI level, and fail to meet the full spectrum of diagnostic needs such as the classification of the return on investment, discovery, retail, and the classification of a full sliding image (WSI) and VQA in clinical practice. In this study, we offer Smartpath-R1, multi-use MLLM capable of treating both ROI and WSI tasks while showing strong pathological thinking capacity. The framework of our work combines the light control based on the scale and the reinforcement based on tasks, which spoils the requirements of the Supervision Series of Thought by taking advantage of the fundamental knowledge within MLLM. Moreover, Smartpath-R1 merges a multi-tasking analysis through the mechanism of experience in a mixture, allowing the dynamic treatment of various tasks. We sponsor a wide range of data collection that includes 2.3 million samples of the return on investment and 188k WSI sample for training and evaluation. Intensive experiences across 72 tasks to verify the validity of the effectiveness and superiority of the proposed approach. This action represents an important step towards developing multi -use artificial intelligence systems and logic of microscopic pathology.