M3Net clinical-inspired 3D network improves lung nodule classification accuracy by 3.26%

A team led by Jinyue Li has introduced M3Net, a novel 3D convolutional network for pulmonary nodule classification that explicitly mimics a radiologist's hierarchical diagnostic approach. Traditional deep learning models for CT analysis often act as black boxes, limiting clinical trust. M3Net addresses this by constructing progressive multi-scale inputs—from fine-grained nodule structures (micro) to local tissue semantics (meso) and global anatomical relationships (macro). Each scale has its own encoder, and the model enforces cross-scale semantic consistency through latent space projection and mutual information maximization, making the decision process more interpretable.

Tested on two datasets, M3Net achieved 86.96% accuracy on the public LIDC-IDRI benchmark and 84.24% on the self-collected USTC-FHLN clinical dataset, outperforming the best existing baselines by 3.26% and 2.17% respectively. The code is open-sourced, and the paper appears in Information Fusion (2026). By aligning model architecture with clinical reasoning, M3Net demonstrates that explainable AI can deliver state-of-the-art performance in critical medical imaging tasks like early lung cancer screening.