Recently, a research team from the Hefei Institutes of Physical Science, Chinese Academy of Sciences, has developed a novel AI medical diagnostic system called MultiXpert. This system enables intelligent chest X-ray diagnosis under "zero-shot" conditions—meaning it requires no labeled training data and can even identify diseases it has never encountered before. This advancement brings AI closer to the diagnostic reasoning patterns of human physicians. Chest X-rays are a widely used imaging method in clinical practice, but their interpretation is time-consuming and highly dependent on expert experience. While traditional AI systems have achieved performance comparable to radiologists in certain tasks, they rely heavily on large volumes of manually annotated data. This dependency limits their ability to generalize across new diseases or variations in data from different hospitals, hindering their effectiveness in complex real-world clinical settings. To address these challenges, the research team introduced a new multi-modal dual-stream collaborative enhancement approach. The resulting system, MultiXpert, achieves high-precision diagnosis without requiring additional labeled data, enabling true zero-shot learning. It simultaneously processes both image and textual information, integrating large language models with expert radiological knowledge to enhance the description of lesions, thereby achieving deep fusion between visual and linguistic understanding. In the image processing branch, MultiXpert employs a lesion-aware masking mechanism to improve feature representation of potential abnormalities without explicit annotations. A hierarchical memory matrix dynamically balances global anatomical context with local lesion features, enhancing the model’s structured understanding of complex medical images. In the text branch, the system combines large language models with clinical expert knowledge to perform semantic calibration and structural reconstruction of disease descriptions, generating reports that are both anatomically precise and clinically standardized, thus improving consistency in medical language expression. A cross-modal semantic alignment module enables complementary fusion of information at multiple granularities, significantly boosting the model’s performance in identifying lesions and making diagnoses under zero-shot conditions. Experimental results show that MultiXpert achieved an average AUC improvement of 7.5% across four single-label public datasets, and outperformed state-of-the-art vision-language models by an average of 3.9% in zero-shot scenarios. This breakthrough offers a new technical pathway for zero-shot chest X-ray diagnosis and represents a paradigm shift in medical AI—from reliance on labeled data to autonomous understanding. The findings have been published in Information Processing and Management, and the work was supported by the National Natural Science Foundation of China. The research provides a significant step forward in making AI more adaptable, interpretable, and clinically applicable in real-world healthcare environments.