CHM-Net beats MRI benchmarks with 12% accuracy gain for tumor microbial density

A team of researchers led by Jiaming Liang and Haolin Chen from several Chinese universities and hospitals has developed CHM-Net (Center Heatmap-driven Macro-Micro Modeling Network), a novel deep learning approach that can predict microbial density stratification directly from multimodal MRI scans without invasive procedures. The work, published on arXiv, addresses the clinically important task of MRI-based Microbial Density Stratification (MRI-MDS). CHM-Net first uses center heatmap-guided small-lesion response localization to establish a link between imaging phenotypes and microbial states, then constructs patient-level macro-micro evidence from these localized heatmap responses for microbial density prediction.

On the newly constructed GBNPC 2026 dataset, CHM-Net outperformed all representative baselines, achieving a 12.06% absolute accuracy improvement over the next best competitor. The model also demonstrated robustness in auxiliary tests on two additional 3D medical image datasets, suggesting its generalizability across volumetric medical image classification scenarios. While the paper focuses on MRI-based inference, the technique could reduce the need for tissue biopsies and enable more frequent, non-invasive tumor monitoring. The code and data are available via the project's GitHub repository.