Shenzhen Institutes of Advanced Technology Unveils AFLoc, an AI Model That Locates Lesions in Medical Images Without Manual Annotation

A research team led by Professor Wang Shanshan of the State Key Laboratory of Biomedical Imaging Science and System at the Shenzhen Institutes of Advanced Technology (SIAT), Chinese Academy of Sciences, published AFLoc, an AI model that finds lesions in medical images without doctors first annotating them, in Nature Biomedical Engineering on January 6, 2026, the Shenzhen Special Zone Daily reported. The work was done with Assistant Professor Zhou Hongyu of Tsinghua University and Professor Zhang Kang of the Institute for AI in Medicine at Macau University of Science and Technology. Training medical imaging AI has usually depended on doctors manually outlining lesions, which is time-consuming and has held back large-scale use. AFLoc instead learns from two kinds of information at once: the images themselves, such as chest X-rays, fundus photographs and pathology slides, and the clinical reports doctors write about them. Through repeated contrastive learning it works out which image regions correspond to the diseases mentioned in the reports, so it can mark the most likely lesion locations without manual labels. The team validated AFLoc on chest X-rays, fundus images and histopathology. In chest X-ray tests covering 34 common chest diseases, including pneumonia, pleural effusion and pneumothorax, across 8 public datasets, it outperformed existing methods on several localisation metrics and matched or exceeded human experts for several diseases. It also localised lesions more accurately than mainstream models on fundus and pathology images, and in zero-shot classification across all three modalities it performed better overall than existing methods; for retinal disease it even beat some models fine-tuned on manually annotated data. "This model effectively avoids traditional deep learning methods' dependence on large-scale manually annotated data," said Wang Shanshan, adding that it significantly improves the use of medical imaging data and the model's ability to generalise. Wang and Zhang are co-corresponding authors, and SIAT PhD student Yang Hao and Zhou Hongyu are co-first authors. The team plans multi-centre validation in real clinical settings to move AFLoc toward clinical decision support.