MBNM: Multi-branch network based on memory features for long-tailed medical image recognition

Area of Science:

• Computer-aided diagnosis
• Deep learning for medical imaging
• Imbalanced data learning

Background:

• Deep learning models in computer-aided diagnosis require balanced datasets.
• Clinical medical data often exhibits imbalanced, long-tailed distributions due to rare diseases.
• Existing methods for imbalanced data increase complexity, inference time, and reduce performance stability.

Purpose of the Study:

• To develop an effective computer-aided diagnosis algorithm for long-tailed medical data.
• To improve the generalization ability of deep learning models on imbalanced datasets.
• To reduce inference time and maintain stable performance in medical image recognition.

Main Methods:

• Proposed the Multi-Branch Network based on Memory Features (MBNM) for long-tailed medical image recognition.
• MBNM employs three branches: regular learning for general features, tail learning with a memory module and reverse sampler for rare classes, and a fusion balance branch with adaptive loss for performance re-balancing.
• The memory module is deactivated during inference to optimize speed.

Main Results:

• Achieved significant improvements in Accuracy, MCR, F1-Score, Precision, and AUC on imbalanced Ophthalmic OCT and Skin-7 datasets.
• Demonstrated superior performance compared to strong baselines in auxiliary diagnosis scenarios with extreme data imbalance.
• Outperformed state-of-the-art models on natural image datasets (CIFAR-10, CIFAR-100).

Conclusions:

• The MBNM algorithm effectively addresses imbalanced data distribution with minimal added cost.
• The memory module's deactivation during inference preserves computational efficiency.
• MBNM exhibits outstanding performance on both medical and natural images across various imbalance factors.