Chinese Academy of Sciences Unveils AFloc: Self-Supervised AI Model for Medical Imaging Lesion Detection

Title

Chinese Academy of Sciences Unveils AFloc: Self-Supervised AI Model for Medical Imaging Lesion Detection

Content

Source: Tencent News, 2026-01-06 In hospitals, a single medical image often hides substantial critical information. But to enable AI to "understand" these images, the process has historically depended on doctors manually circling lesions as training data — an approach that is not only time- and labor-intensive, but has also become a major bottleneck preventing large-scale deployment of medical imaging AI. Is it possible for AI to learn to "find lesions itself" without relying on manual annotation? On January 6, a research team led by Professor Wang Shanshan from the State Key Laboratory of Biomedical Imaging Science and System at the Chinese Academy of Sciences (Shenzhen Institute of Advanced Technology), in collaboration with Assistant Professor Zhou Hongyu of Tsinghua University and Professor Zhang Kang of Macau University of Science and Technology, published a new result in Nature Biomedical Engineering that provides a brand-new answer. The research team proposed an AI model named AFloc, whose defining characteristic is: it does not require doctors to pre-annotate lesions; it can automatically "find lesions" in medical images. In this study, Professor Wang Shanshan and Professor Zhang Kang (Director of the Medical AI Research Institute at Macau University of Science and Technology) served as co-corresponding authors; PhD student Yang Hao from the Imaging Center of SIAT and Assistant Professor Zhou Hongyu of Tsinghua University served as co-first authors. SIAT is the primary completion and corresponding unit. The work was also guided and supported by Academician Zheng Hairong. The traditional learning paradigm for medical imaging AI models is like a student doing exercises who must first have the standard answers. AFloc is more like learning to understand images by "looking at images and reading reports" simultaneously. The researchers explain: "We trained AFloc to learn two types of information simultaneously — one is the medical image itself (e.g., chest X-rays, fundus photos, or pathology slices), and the other is the clinical report written by physicians. Through repeated 'comparative learning,' AFloc gradually understands: when a clinical report mentions a disease, which regions in the image it corresponds to. Over time, even without manual annotation, it can accurately mark the most likely lesion locations in images." The research team systematically validated AFloc on three typical medical imaging modalities — chest X-ray, fundus imaging, and histopathology — and the results showed excellent performance across all of them. In chest X-ray experiments, AFloc was tested on 34 common chest conditions (including pneumonia, pleural effusion, pneumothorax) across 8 mainstream public datasets. It outperformed existing methods on multiple lesion localization metrics, and reached or even surpassed human expert performance on several diseases. In fundus imaging and pathology tasks, AFloc similarly demonstrated stable lesion localization capabilities, with localization accuracy superior to current mainstream models. Beyond lesion localization, AFloc also showed strong disease diagnosis capabilities. In zero-shot classification tasks on chest X-ray, fundus, and histopathology images, its overall performance surpassed existing methods. Particularly in fundus retinopathy diagnosis, AFloc's zero-shot classification performance even exceeded some models that depend on fine-tuning with manually annotated data. "This model effectively avoids the traditional deep learning approach's dependency on large-scale manually annotated data, significantly improving the utilization efficiency of medical imaging data and the model's generalization capability. It provides a feasible path for clinical imaging AI to move from 'relying on manual annotation' to 'self-supervised learning,' and offers a new technical paradigm for building more intelligent and more generalizable medical AI systems," said Professor Wang Shanshan. In the future, the research team will further promote the validation and application of AFloc in multi-center real clinical scenarios, accelerating its translation into clinical auxiliary diagnostic systems. Paper link: https://www.nature.com/articles/s41551-025-01574-7