SJTU, Chuangzhi, and Ruijin Jointly Release CX-Mind: Chest X-ray Diagnosis Enters the Era of 'Verifiable Reasoning'
Shanghai Jiao Tong University, Shanghai Chuangzhi Institute, and Ruijin Hospital have unveiled CX-Mind, the first multimodal large model that transforms chest X-ray diagnosis into a verifiable reasoning chain. It achieves an average 25.1% improvement across three capability domains on 23 datasets and ranks first in all five evaluation dimensions by doctors on a real-world test set.
A consortium led by Shanghai Jiao Tong University, Shanghai Chuangzhi Institute, and Ruijin Hospital has introduced CX-Mind, a multimodal large language model that advances chest X-ray diagnosis from simple classification to a verifiable reasoning chain. Unlike previous AI systems that output only diagnoses or labels, CX-Mind provides a step-by-step reasoning process grounded in imaging evidence, allowing physicians to review, question, and verify each conclusion.
The model was evaluated on a comprehensive benchmark spanning 23 public datasets, totaling 708,473 chest X-ray images and 2.6 million instruction samples. Across three core capability domains—visual understanding, text generation, and spatiotemporal alignment—CX-Mind achieved an average performance improvement of 25.1% over existing specialized models. In the real-world Rui-CXR test set, which includes 4,031 high-quality chest X-rays from Ruijin Hospital covering 14 common thoracic diseases, CX-Mind outperformed all baselines. Furthermore, a multi-center subjective evaluation by clinicians with varying experience levels ranked CX-Mind first in all five dimensions: Clinical Relevance, Logical Coherence, Evidence Support, Differential Diagnostic Coverage, and Explanation Clarity.
CX-Mind introduces three major breakthroughs. First, it adopts an interleaved reasoning paradigm that alternates between thinking steps and answer outputs, making the diagnostic process transparent and auditable. Second, the team constructed CX-Set, a large-scale instruction dataset that covers a complete spectrum of chest X-ray expertise, from disease identification to report generation and spatiotemporal comparison. Third, they developed CuRL-VPR, a curriculum-based reinforcement learning method that rewards not only correct final answers but also the quality of intermediate reasoning steps, verified against real radiology reports.
The implications of CX-Mind extend beyond chest X-rays. Its architecture of verifiable, collaborative reasoning sets a foundation for future medical AI agents that can work alongside doctors across multiple imaging modalities and clinical workflows. While further prospective studies and regulatory approvals are needed, CX-Mind represents a clear step toward AI systems that are not just accurate, but also trustworthy and interpretable in clinical settings.