EMDA-Net: A New Scheme for Medical Image Classification Integrating Earth Mover's Distance and Attention Mechanism

Original Author/Maintainer: SuryaMajumder Source Platform: GitHub Original Project Name: EMDA-Net: Earth Mover's Distance influenced Attention-aided Network for Medical Image Classification Original Link: https://github.com/SuryaMajumder/EMDA-Net-Earth-Mover-s-Distance-influenced-Attention-aided-Network-for-Medical-Image-Classification Publication Date: 2026-05-27

Core View: EMDA-Net organically combines Earth Mover's Distance (EMD) with attention mechanism to provide a new solution for medical image classification tasks, aiming to address key challenges in this field.

Medical image classification faces multiple difficulties:

1. Class imbalance: Normal samples are far more numerous than lesion samples;
2. Small proportion of lesion areas: It is difficult to accurately locate and identify key regions;
3. Data heterogeneity: Images collected from different devices and hospitals have significant differences in brightness, contrast, resolution, etc., requiring models to have strong generalization capabilities.

Introduction of Earth Mover's Distance (EMD)

EMD (Wasserstein distance) measures the difference between two probability distributions, considers spatial structure information, can more accurately describe the similarity of feature distributions, and is suitable for the complex feature distributions of medical images.

EMD-Influenced Attention Mechanism

The attention mechanism allows the model to focus on key regions (e.g., lesions). The attention module of EMDA-Net is influenced by EMD, which can dynamically adjust the focus according to differences in feature distributions and understand the diagnostic importance of different regions.

EMDA-Net adopts an encoder-decoder paradigm:

1. Convolutional layers extract multi-scale features;
2. EMDA attention modules are applied at different levels (calculating the EMD between query feature and reference feature distributions, converting to attention weights, and weighted aggregation of features);
3. Fully connected layers complete the classification.

1. Efficient EMD calculation: Using approximate algorithms or differentiable EMD variants to reduce computational complexity;
2. End-to-end training: Carefully designed loss functions and optimization strategies, possibly using multi-task learning and progressive training to improve convergence and performance.

EMDA-Net provides a new path for intelligent analysis of medical images:

• Compared with traditional CNNs and ordinary attention models, it can better capture subtle lesion features and help with early disease screening;
• Assists radiologists in improving diagnostic efficiency and accuracy, reducing missed diagnoses and misdiagnoses;
• Provides a reliable preliminary screening tool for areas with scarce medical resources.

Limitations

• The computational complexity of EMD may limit its application in ultra-high-resolution images;
• The interpretability of the attention mechanism needs to be further aligned with clinical knowledge.

Future Directions

• Extend to 3D medical images (CT, MRI volume data);
• Combine multi-modal information (images + clinical data) for comprehensive diagnosis;
• Explore more efficient EMD approximation algorithms to reduce computational overhead.