Addressing Class Imbalance with Latent Diffusion-based Data Augmentation for Improving Disease Classification in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) models struggle with imbalanced datasets common in medical imaging.
• Synthetic data augmentation can improve DL model performance and generalization.
• Latent diffusion models (LDMs) show potential for high-quality medical image synthesis.

Purpose of the Study:

• To evaluate the effectiveness of text-guided image-to-image LDMs for synthesizing disease-positive chest X-rays (CXRs).
• To augment a pediatric CXR dataset with synthesized images to improve classification performance.
• To mitigate class imbalance and enhance generalization in medical image classification tasks.

Main Methods:

• Established baseline performance of an Inception-V3 model on imbalanced CXR data (normal vs. pneumonia/bronchopneumonia).
• Fine-tuned text-guided LDMs to generate synthetic CXRs depicting pneumonia and bronchopneumonia.
• Retrained the Inception-V3 model using augmented datasets including LDM-synthesized images.

Main Results:

• LDM-synthesized image augmentation significantly improved Youden's index (p<0.05).
• Augmentation markedly enhanced other classification metrics including balanced accuracy, sensitivity, specificity, F-score, MCC, and Kappa.
• The strategy effectively addressed class imbalance and improved model generalization.

Conclusions:

• Text-guided LDM-based data augmentation is a viable strategy for improving deep learning classification on imbalanced medical imaging datasets.
• Synthesizing disease-positive CXRs with LDMs can enhance model robustness and diagnostic accuracy.
• This approach offers a promising solution for overcoming data limitations in medical AI research.