EXPLAINABLE MULTIMODAL GENERATIVE AI FRAMEWORK FOR DISEASE DETECTION USING MEDICAL IMAGES AND CLINICAL REPORTS

Abstract

In the context of medicine, recent advancements in generative Artificial Intelligence (GenAI)-especially the maturity of Generative Adversarial Networks (GANs), Varia tional Autoencoders (VAEs), and the advent of progressive denoising diffusion models have significantly transformed the way clinical practitioners assess medical images and diagnose serious medical conditions. This thesis is the first to deeply cross-domained study the generative paradigms on five crucial pathological domains: neurology degeneration (Alzheimer’s disease), lungs diseases (pneumonia, COVID-19, and tuberculosis), ocular maladies (glaucoma), orthopedic diseases (knee osteoarthritis), and hyper contagious skin diseases (monkeypox). We evaluate twenty-five state-of the-art generative models based on their achievement of specific medical tasks, the reliance on particular clinical datasets, image realism and structure accuracy trade-offs across public benchmarks (ADNI, OAI, NIH Chest X-Ray, and MSID) on these five pathological domainspublishedfrom2020 to2026,highlightinghowgenerativemodels overcome practical issues such as data scarcity, severe class imbalance and stringent patient privacy. Three important translational aspects are studied in this research to link laboratory success to actual clinical practice: the mapping of Explainable AI (XAI) frameworks (Grad-CAM and Med-CAM) onto generated images to boost physicians’ trust in the diagnosis; privacy-preserving decentralized machine learning (Federated Learning using Flower) to collaborate multi-centers safely without disclosing sensitive patient scans; and global digital health regulatory procedures (FDA 510(k), De Novo tracks and Predetermined Change Control Plans). This systematic study constitutes an engineering and regulatory roadmap to pave the way for safe and ethical integration of GenAI in robust diagnostic support systems.