MUSCLE: A New Perspective to Multi-Scale Fusion for Medical Image Classification Based on the Theory of Evidence

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Medical image classification is crucial but challenged by lesion variability.
• Existing computer vision backbones struggle with multi-scale features in medical images.
• Single-scale feature extraction limits accuracy for scattered and size-varying lesions.

Purpose of the Study:

• To propose a novel multi-scale learning paradigm, MUSCLE, for improved medical image classification.
• To address limitations of single-scale feature extraction in medical image analysis.
• To enhance feature representation for accurate classification of diverse medical images.

Main Methods:

• Developed MUlti-SCale Learning with trusted Evidences (MUSCLE) paradigm.
• Utilized evidence theory to estimate feature uncertainties and trustworthiness.
• Employed evidential deep neural networks to form opinions on multi-scale features.
• Ensembled opinions to adaptively tune feature weights for improved classification.

Main Results:

• MUSCLE improved accuracy on five public medical image datasets.
• The paradigm enhanced the reliability and interpretability of classification models.
• Experimental results demonstrate superior performance compared to standard backbones.

Conclusions:

• MUSCLE offers a robust solution for medical image classification challenges.
• The integration of multi-scale features with evidence theory boosts model performance.
• The method provides trusted evidence for more reliable AI-driven medical diagnoses.