Medical imaging datasets usually have a serious problem: class imbalance. For rare diseases like pneumonia, the numbers are especially skewed—pneumonia might only show up in 2-5% of the cases. This messes with machine learning models, making them lean way too heavily on the majority class. The result? They miss actual disease cases and can’t be trusted for clinical predictions.

 

So, for this project, I built a new, practical solution: using class-conditional diffusion probabilistic models to generate synthetic medical images, specifically to tackle rare event detection.

 

Here’s how I went about it. First, I created a synthetic chest X-ray generator that actually looks and behaves like the real thing. It simulates realistic lung fields, the heart outline, ribs, and the spine—all with noise that mimics what you get in actual X-rays. On top of that base, I trained a class-conditional CNN-based diffusion model with 100 timesteps to create high-quality fake pneumonia chest X-rays. These images aren’t generic—they show true-to-life consolidation, opacity gradients, and bronchial markings.

 

Using this method, I generated 500 new synthetic pneumonia samples. This had a huge impact on the original data imbalance. To put it in perspective, the original dataset was horribly skewed: just 113 pneumonia cases out of 3,875 normal ones (that’s a ratio of 34.3:1). After adding the synthetic images, pneumonia cases went up to 613, and the imbalance dropped to 6.3:1—an 81.6% improvement. That’s a 445% increase in pneumonia samples.

 

Next, I trained a CNN classifier on this newly balanced dataset and tested it on the original skewed test set. The results were kind of wild—100% across the board. Sensitivity, specificity, precision, F1-score, accuracy—all perfect. The confusion matrix showed not a single false positive or false negative. The model completely nailed the task of telling normal and pneumonia cases apart.

This project is clear proof that using diffusion-based synthetic data can actually fix extreme class imbalance in medical imaging, making rare event detection trustworthy for clinical use. It lays the groundwork for more research into synthetic medical data, privacy-preserving AI, and rare disease detection in other areas, too. The whole system—from the anatomy simulator to the diffusion model to the classifier—is fully end-to-end and reproducible. In other words, it’s a solid framework for anyone struggling with class imbalance in healthcare AI.

Keywords— Diffusion Models; Class-Conditional Generation; Rare Event Detection; Extreme Class Imbalance; Medical Imaging; Chest X-Ray; Pneumonia Detection; Synthetic Data Generation; Deep Learning; Convolutional Neural Networks; DDPM; Generative AI; Healthcare AI; Medical Image Analysis.