VariMix: A variety-guided data mixing framework for explainable medical image classifications.

Area of Science:

• Computer Science
• Artificial Intelligence
• Medical Imaging

Background:

• Deep neural networks require data augmentation to prevent overfitting and improve generalization.
• Generative Adversarial Networks (GANs) synthesize realistic images but often lack diversity and have ambiguous labels.
• Current data mixing strategies based on salient regions may not capture all diagnostic information, leading to label mismatches.

Purpose of the Study:

• To address label mismatches in medical image classification caused by data augmentation techniques.
• To propose a novel data mixing framework that improves the accuracy and diversity of synthetic medical images.

Main Methods:

• Introduced VariMix, a variety-guided data mixing framework.
• Employed an absolute difference map (ADM) generated by an image-to-image (I2I) GAN to resolve label mismatches.
• Enabled bidirectional mixing operations between training samples for enhanced data synthesis.

Main Results:

• VariMix achieved high accuracies: 99.30% (CXR) and 94.60% (Retinal) with SwinT V2.
• ConvNeXt classifier achieved top accuracies: 87.73% (Breast US), 99.28% (CXR), 95.13% (Retinal), and 95.81% (Maternal-Fetal US).
• Medical expert evaluation confirmed the I2I GAN's potential to improve medical image classification accuracy.

Conclusions:

• VariMix outperforms existing GAN- and Mixup-based methods on four public datasets.
• The I2I GAN provides interpretability for medical image classifications by generating hyperplane difference maps.
• The proposed method offers a superior approach for data augmentation in medical deep learning.