Weighted Sampling with Frequency-Aware Spatial Attention for Imbalanced Image Classification

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Class imbalance is a major challenge in image classification, leading to poor performance on underrepresented classes.
• Existing methods struggle to adequately address severe class imbalance in medical datasets.

Purpose of the Study:

• To propose a hybrid framework (WSFSA) that addresses class imbalance at both data and feature levels.
• To enhance the recognition performance of minority classes in image classification tasks.

Main Methods:

• Developed a hybrid framework combining weighted sampling and frequency-aware spatial attention (WSFSA).
• Weighted sampler increases training exposure for minority classes.
• Frequency-aware spatial attention module integrates class frequency information to boost feature discrimination for underrepresented classes.

Main Results:

• WSFSA demonstrated significant benefits on severely imbalanced datasets (DermaMNIST, DermaMNIST-224), performing comparably to top baseline methods.
• Moderate gains were observed on OrganCMNIST, with similar performance across methods on the less imbalanced BloodMNIST.
• Per-class analysis confirmed improved sensitivity for minority/difficult classes while maintaining high specificity.

Conclusions:

• Combining sampling-level and feature-level rebalancing strategies is a practical approach for improving class-balanced recognition.
• The WSFSA framework shows particular effectiveness under conditions of severe class imbalance.
• This hybrid method offers a robust solution for enhancing model performance on imbalanced medical image datasets.