Article · AI translation
IT Home, November 13: A research team from the Institute of Health and Medicine at the Hefei Institutes of Physical Science, Chinese Academy of Sciences, and Hefei Cancer Hospital recently worked with several institutions in China and abroad to make progress in lung cancer diagnosis research. They developed a new AI-assisted method for diagnosing pulmonary nodules based on chest CT images, and published their findings in Information Fusion.
Lung cancer is a common malignant tumor. Early screening is key to improving cure rates and reducing mortality, while accurately determining whether pulmonary nodules found during early screening are benign or malignant directly affects subsequent treatment plans and outcomes.
In clinical practice, doctors must examine chest CT images one by one, considering the overall structure of the lungs, tissue surrounding a nodule, and details of the nodule, then draw on clinical experience to determine whether it is benign or malignant. This traditional manual image review is time-consuming and labor-intensive, is susceptible to factors such as experience and fatigue, and can even lead to missed or incorrect diagnoses.
AI technology that has emerged in recent years has been used to assist in determining whether pulmonary nodules are benign or malignant. However, conventional AI diagnostic models make judgments based only on image features at a single scale and are mostly “black box” algorithms. Their limited accuracy and poor interpretability make it difficult for them to meet clinical needs for precise diagnosis and treatment.
To address these problems, the research team designed an algorithm that imitates doctors’ diagnostic reasoning. They proposed a new AI model, M³Net, a “macro-meso-micro” hierarchical 3D neural network, and developed a new multiscale, interpretable method for intelligent classification and diagnosis of pulmonary nodules.
The algorithm uses three distinct image input scales corresponding to the overall macroscopic structure of the lungs, features of the tissue surrounding a nodule, and microscopic details of the nodule, comprehensively covering important information relevant to diagnosing the lesion.
The model uses dedicated encoders to extract image features at each scale and aligns features across scales through latent-space projection and mutual information maximization. Using a hierarchical cross-attention mechanism, it simulates doctors’ diagnostic process of looking first at the whole, then the local area, and finally the details. After optimizing feature fusion through multiple regularization constraints, it captures subtle signs of malignancy such as nodule lobulation, spiculation, and uneven density while also accounting for broader pathological information such as vascular traction and deformation of lung tissue, enabling a detailed, three-dimensional analysis of lesions.
To test the model’s effectiveness, the research team conducted comparative experiments using the public LIDC-IDRI dataset and the clinical USTC-FHLN dataset. The results showed that, compared with 12 popular AI algorithms, M³Net improved accuracy and other measures in classifying pulmonary nodules as benign or malignant in both datasets. It had a particularly clear advantage in distinguishing difficult nodules with indistinct borders and atypical features, while also requiring fewer model parameters and less computational cost.
The multiscale image-input AI algorithm proposed in this study follows doctors’ diagnostic reasoning, making it easier to establish effective validation between humans and machines and offering a new strategy for improving the accuracy and relative interpretability of AI-based lung cancer diagnosis.
The Hefei Institutes of Physical Science, Chinese Academy of Sciences, was the primary institution responsible for this work. Collaborating institutions included the First Affiliated Hospital of the University of Science and Technology of China (Anhui Provincial Hospital), Bengbu Medical University, the Institute of Semiconductors of the Chinese Academy of Sciences, Northeastern University in the United States, and Nanyang Technological University in Singapore. The study received support from Hefei Cancer Hospital of the Chinese Academy of Sciences’ single-disease research project, among other sources.