Handling Imbalanced Data: Uncertainty-Guided Virtual Adversarial Training With Batch Nuclear-Norm Optimization for...

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Medical image datasets often exhibit class imbalance, leading to biased model predictions.
• Semi-supervised learning (SSL) exacerbates this issue by generating pseudo-labels from biased models.
• This bias hinders the effective utilization of both labeled and unlabeled medical image data.

Purpose of the Study:

• To propose a novel semi-supervised deep learning method for large-scale medical image classification.
• To address the challenges posed by imbalanced datasets in medical image analysis.
• To enhance the discriminability, diversity, and generalization of models trained on medical images.

Main Methods:

• Developed an uncertainty-guided virtual adversarial training (VAT) method combined with batch nuclear-norm (BNN) optimization.
• Employed a multi-loss strategy including supervised cross-entropy, lBNN loss, uBNN loss, and VAT loss.
• Integrated uncertainty estimation to filter unlabeled samples near the decision boundary for VAT loss computation.

Main Results:

• The proposed method demonstrated superior performance compared to existing state-of-the-art SSL techniques.
• Experiments were conducted on two public and one in-house medical image datasets.
• The approach effectively leveraged information from both labeled and unlabeled data to improve classification accuracy.

Conclusions:

• The uncertainty-guided VAT with BNN optimization offers a robust solution for imbalanced medical image classification.
• This method effectively mitigates bias issues inherent in semi-supervised learning on imbalanced data.
• The proposed technique shows significant potential for improving diagnostic accuracy in clinical settings.