GENERATIVE DIFFUSION MODELS FOR ANOMALY DETECTION IN HEALTHCARE

Abstract

Chest X-rays are one of the most common and vital diagnostic tools for detecting a variety of pulmonary and cardiac conditions. However, the interpretation of these images often relies on the expertise of radiologists, and can be subject to human error, particularly in high-workload or low-resource settings. This thesis introduces a novel application of diffusion models, a type of generative artificial intelligence, for the automated detection of anomalies in chest X-rays. Unlike traditional convolutional neural networks, diffusion models generate new samples by gradually denoising a random signal and have shown promising results in maintaining high fidelity in image generation tasks. Here, we adapted diffusion models to identify deviations from typical X-ray images, effectively spotting anomalies such as tumors, fractures, and pneumothoraces. Our methodology involved training the diffusion model on a large dataset of normal chest X-rays, allowing the model to learn a distribution of common patterns and features in healthy patients. Anomalies were then detected by the model’s inability to reproduce these features accurately in diseased states, which was quantified using a novel anomaly scoring system. The performance of our model was evaluated against a benchmark dataset and compared to traditional anomaly detection techniques. The results demonstrate that diffusion models not only enhance the accuracy of anomaly detection in chest X-rays but also reduce false positives and increase the interpretability of the findings. This research contributes to the field by providing a robust framework for medical imaging analysis that can support radiologists in diagnostic processes and potentially improve patient outcomes through early detection.