Chinese researchers develop multi-scale AI model mimicking physician workflow for lung nodule diagnosis

A research team from the Health Institute of Hefei Institute of Materials, Chinese Academy of Sciences, in collaboration with various domestic and international partners, has advanced lung cancer diagnosis by creating a novel AI-assisted method for analyzing chest CT images. This new approach, detailed in the journal Information Fusion, addresses the limitations of traditional manual reading, which is time-consuming and susceptible to human error. The developed system, named M³Net, is a hierarchical 3D neural network designed to emulate a physician's diagnostic thought process—moving from global to local to fine details. It processes three distinct image scales corresponding to the lung's macro structure, nodule periphery, and microscopic features, using cross-scale attention mechanisms for comprehensive analysis. Testing on the LIDC-IDRI and USTC-FHLN datasets demonstrated that M³Net outperformed twelve existing AI algorithms in classification accuracy, showing particular strength in distinguishing ambiguous nodules while maintaining lower computational costs. The research, supported by projects from the Hefei Tumor Hospital, involves collaborations with institutions such as Northeastern University and Nanyang Technological University.