Enhancing Biomedical Imaging with AI: Compression, Prediction, and Multi-Modal Integration for Clinical Advancement

Abstract

This dissertation delves into the enhancement of biomedical image analysis through the deployment of artificial intelligence methodologies, focusing on the transition from theoretical innovation to practical clinical utility. Spanning four cornerstone projects, the work encapsulates the development of predictive models for spatial transcriptomics, efficient image compression for cancer pathology slides, and critical evaluations of histopathology slide search engines. The first project employs Random Forest Regression and spatial point processes to forecast cell distribution patterns, thereby offering a novel perspective on gene expression in embryogenesis at a single-molecule resolution. The second venture introduces a Variational Autoencoder (VAE) that sets a new precedent in histopathology imaging with a significant compression ratio, maintaining diagnostic reliability. Lastly, the third project assesses the performance of leading histopathology slide search engines, establishing a benchmark for their clinical application and suggesting enhancements for future integration. Together, these projects pave the way for AI-driven approaches to be woven into the fabric of clinical practice, signaling a transformative leap in the utility of biomedical imaging and multi-channel data interpretation