M-Net: Integrating Spectral Features and Physical Field Operators into Deep Learning for Medical Image Segmentation

Research Question and Approach

Medical image segmentation often relies on purely data-driven deep learning models, which may overlook the inherent geometric and physical structure of anatomical data. The authors propose M-Net (Math-Augmented Network), a framework that explicitly incorporates mathematical priors into the U-Net architecture. The approach integrates three specific mathematical components: a continuous spectral feature derived from the condition number of local pixel matrices, physical field operators (divergence and a curl-like operator) to capture boundary and intensity characteristics, and a Math-Attention Gate (MAG) to adaptively fuse these priors with deep features.

Methodology

The core innovation lies in the use of the condition number of mean-centered local pixel matrices. By mean-centering the 3x3 neighborhood, the authors create a differentiable, scale-invariant measure of local texture complexity that avoids the pitfalls of naive intensity-based calculations. Additionally, the network computes divergence and a discrete curl-like operator from image gradients to highlight focal intensity extrema and boundary non-smoothness. These mathematical features are not simply concatenated; they are processed through the MAG, which uses the mathematical priors to generate spatial weight masks, allowing the network to selectively emphasize regions with anomalous mathematical properties.

Experiments on the LiTS (liver), KiTS (kidney), and BraTS (brain tumor) datasets show that M-Net consistently outperforms the baseline U-Net. The model achieved Dice scores of 78.42%, 76.15%, and 83.67% respectively, representing significant improvements over the baseline. Ablation studies confirm that the continuous condition-number feature provides a 2.14% performance boost over binary invertibility methods, and the MAG mechanism adds a 1.45% gain compared to simple feature concatenation, validating the importance of the proposed fusion strategy.