Toward Multicenter Skin Lesion Classification Using Deep Neural Network With Adaptively Weighted Balance Loss

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks excel at skin lesion recognition from dermoscopic images.
• Existing methods often neglect data imbalance across multi-center clinics, limiting performance.
• Varied data distributions in different clinical centers challenge model flexibility and accuracy.

Purpose of the Study:

• To address the data imbalance issue in multi-center skin lesion classification.
• To propose a novel adaptively weighted balance (AWB) loss function.
• To enhance the flexibility and accuracy of classification networks in diverse clinical settings.

Main Methods:

• Introduction of a novel adaptively weighted balance (AWB) loss to conventional classification networks.
• Focus shifted from network framework improvement to addressing data imbalance.
• AWB loss enables adaptive intraclass compactness and prioritizes minority classes.

Main Results:

• The proposed AWB loss solution demonstrates superior flexibility and competence in multi-center imbalanced skin lesion classification.
• Achieved considerable performance on two benchmark datasets compared to state-of-the-art loss functions.
• Effectiveness validated in imbalanced gastrointestinal disease classification and DR grading tasks.

Conclusions:

• The AWB loss is a user-friendly, hyperparameter-free solution for imbalanced multi-center classification tasks.
• The method enhances model adaptability to different clinical data distributions.
• This approach offers a promising direction for improving AI in medical diagnostics.