Addressing Small and Imbalanced Medical Image Datasets Using Generative Models: A Comparative Study of DDPM and PGGANs with Random and Greedy K Sampling

TL;DR

This study addresses the challenge of small, imbalanced medical image datasets under privacy constraints by evaluating synthetic data from DDPM and PGGAN for augmentation. It compares two sampling schemes, Random and Greedy K, across four classifiers, revealing that DDPM consistently achieves lower Fréchet Inception Distance and yields stronger improvements in accuracy and stability, especially on imbalanced data. Random Sampling provides more stable gains, while Greedy K Sampling offers greater diversity at the cost of fidelity. The findings support using diffusion-based augmentation to balance and enlarge medical image datasets, enhancing classifier robustness in clinically relevant settings.

Abstract

The development of accurate medical image classification models is often constrained by privacy concerns and data scarcity for certain conditions, leading to small and imbalanced datasets. To address these limitations, this study explores the use of generative models, such as Denoising Diffusion Probabilistic Models (DDPM) and Progressive Growing Generative Adversarial Networks (PGGANs), for dataset augmentation. The research introduces a framework to assess the impact of synthetic images generated by DDPM and PGGANs on the performance of four models: a custom CNN, Untrained VGG16, Pretrained VGG16, and Pretrained ResNet50. Experiments were conducted using Random Sampling and Greedy K Sampling to create small, imbalanced datasets. The synthetic images were evaluated using Frechet Inception Distance (FID) and compared to original datasets through classification metrics. The results show that DDPM consistently generated more realistic images with lower FID scores and significantly outperformed PGGANs in improving classification metrics across all models and datasets. Incorporating DDPM-generated images into the original datasets increased accuracy by up to 6%, enhancing model robustness and stability, particularly in imbalanced scenarios. Random Sampling demonstrated superior stability, while Greedy K Sampling offered diversity at the cost of higher FID scores. This study highlights the efficacy of DDPM in augmenting small, imbalanced medical image datasets, improving model performance by balancing the dataset and expanding its size.

Figures (8)

• Figure 1: Implemented comprehensive framework that rigorously evaluates the quality and performance of synthetic images generated by DDPM and PGGAN models.
• Figure 2: PGGANs gradually increase both the resolution of the generated images and the complexity of the generator and discriminator networks during training.
• Figure 3: The directed graphical model of DDPM. The forward process gradually adds noise to an image until it becomes pure noise. The backward process then learns to reverse this process, reconstructing the original image from noise.
• Figure 4: PGGANs model training loss for Pneumonia class. Plot indicates the training process is stable where loss function increases as model adds new layers.
• Figure 5: Comparison between Original vs. DDPM vs. PGGANs 160K Pneumonia class images. Synthetic images exhibit higher similarity to the originals.