Abstract
AI education and research: examples of proof of concept or AI in development, technical advances, teaching approaches or preclinical testingData scarcity remains a prevalent issue in the development of artificial intelligence (AI)-based diagnostic tools.While large data sets are usually available for common conditions, rare diseases often lack sufficient samples, leading to a diminished performance and applicability of machine learning models.Generative models provide a promising solution by synthesising data to mitigate class imbalance and improve classifier performance and robustness.We introduce a novel diffusion-based framework designed for data augmentation in low-resource medical scenarios.Our approach extends a basic denoising diffusion probabilistic model through a multi-step class-aware training pipeline, additional attention mechanisms and a GAN-inspired fine-tuning process.This combination is designed to enhance the level of detail in the images, encourage realism and increase the model's ability to learn class-specific details in resource-scarce scenarios.We evaluate this method across diverse medical imaging modalities, including optical coherence tomography (OCT) and cervical cytology images.The synthetic images are assessed using metrics such as CLIP score, and they are evaluated in a downstream classification task.We measure the accuracy of a ResNet-18 model trained on the unaugmented and augmented data sets, evaluated on the same set of real, unseen data.Across OCT experiments, we observed an increase of classifier accuracy from 83% to 92%, with the synthesised images achieving a CLIP score of 0.96.Preliminary results for cervical cytology further demonstrate an increase of diagnostic accuracy from 56% to 62%, and CLIP score of 0.94.Our results show that the model can generate realistic, class-specific images even in data sets containing less than 1,500 images, and classes with as few as 22 samples.These findings highlight the potential of diffusion-based data augmentation as a powerful tool for improved diagnostic performance of under-represented diseases within machine learning.