Newshot Chinese Scientists Propose New AI-Assisted Diagnosis Method for Pulmonary Nodules: Simulating Doctors' Image-Reading Workflow for More Accurate Judgment
IT之家, November 13: A research team from the Health Research Institute and Hefei Cancer Hospital of the Hefei Institutes of Physical Science, Chinese Academy of Sciences, together with multiple domestic and international institutions, has recently made progress in lung cancer diagnosis research, developing a new AI-assisted diagnosis method for pulmonary nodules based on chest CT images. The results were published in Information Fusion.
Lung cancer is a highly prevalent malignant tumor. Early screening is a key step in improving cure rates and reducing mortality, and accurate judgment of the benign or malignant nature of pulmonary nodules in early screening is directly related to subsequent treatment planning and prognosis.
In clinical practice, doctors need to read chest CT images one by one, integrate the overall structure of the lungs, the tissues surrounding the nodule, and nodule details, and then combine clinical experience to diagnose whether a pulmonary nodule is benign or malignant. This traditional manual image reading is not only time-consuming and labor-intensive but also susceptible to factors such as experience and fatigue, and may even lead to missed diagnoses or misdiagnoses.
In recent years, emerging AI technology has been used for AI-assisted diagnosis of benign and malignant pulmonary nodules. However, conventional AI diagnostic models rely only on single-scale image features to make judgments and are mostly “black-box” algorithms, with limited recognition accuracy and poor interpretability, making it difficult to meet the clinical needs of precision diagnosis and treatment.
To address these problems, the research team imitated doctors' diagnostic logic in algorithm design and proposed a new M³Net, a “macro-meso-micro” hierarchical 3D neural network AI model, developing a multi-scale, interpretable new method for intelligent classification and diagnosis of pulmonary nodules.
The algorithm constructs three differentiated image input scales, corresponding respectively to the global macroscopic structure of the lungs, the features of tissues surrounding the nodule, and the microscopic details of the nodule, comprehensively covering important information relevant to lesion diagnosis.
The model uses dedicated encoders to extract image features at different scales, relies on latent space projection and mutual information maximization techniques to align cross-scale features, and uses a hierarchical cross-attention mechanism to simulate doctors' diagnostic workflow of “first global, then local, and finally details.” Through multiple regularization constraints to optimize feature fusion, it captures malignant subtle features such as nodule lobulation, spiculation, and heterogeneous density, while also taking into account global pathological information such as vascular traction and lung tissue deformation, achieving refined and three-dimensional analysis of lesions.
To test the effectiveness of the model, the research team conducted algorithm comparison experiments based on the public dataset LIDC-IDRI and the clinical dataset USTC-FHLN. The results showed that, compared with twelve popular AI algorithms, the M³Net model improved indicators such as accuracy in classifying benign and malignant pulmonary nodules in both datasets. It showed particularly obvious advantages in distinguishing difficult pulmonary nodules with unclear boundaries and atypical features, and the model has lower parameter count and computational cost.
The multi-scale image input AI algorithm proposed in this study, which follows doctors' diagnostic logic, facilitates effective validation between humans and machines, and provides a new strategy for improving the accuracy and relative interpretability of AI diagnosis of lung cancer.
The first completing institution of the achievement is the Hefei Institutes of Physical Science, Chinese Academy of Sciences. Collaborating institutions include The First Affiliated Hospital of University of Science and Technology of China (Anhui Provincial Hospital), Bengbu Medical University, the Institute of Semiconductors, Chinese Academy of Sciences, Northeastern University in the United States, Nanyang Technological University in Singapore, and other institutions. This research was supported by the single-disease key project of Hefei Cancer Hospital, Chinese Academy of Sciences, and other projects.
Related articles
- Microsoft launches AI medical diagnosis system MAI-DxO: accuracy reaches 83% for complex cases, results published in Science
- Research team survey finds over 70% of English biomedical papers have used AI-assisted writing, calls for industry to establish sound regulatory mechanisms
- Experts warn that medical students' overreliance on AI may prevent them from developing diagnostic reasoning ability from the very beginning
- British Medical Association: misleading “AI doctors” pose a huge threat to public safety
- Hainan: fully prohibit illegal practices such as AI automatic prescription, fake consultations, and retroactive prescription supplementation
- National Health Commission: explore promoting and applying technologies such as medical AI-assisted diagnosis in urban community health service institutions