Enhancing Medical Image Utility: Dataset Distillation and Endoscopic Panoramic Reconstruction

Abstract

The advancement of medical imaging plays a pivotal role in diagnostics, yet challenges such as data sharing constraints and limited field-of-view hinder its full potential. This thesis explores two critical areas: dataset distillation for medical imaging and endoscopic panoramic reconstruction. Dataset distillation offers a promising avenue for reducing data size while retaining diagnostic efficacy, addressing challenges in data privacy and storage efficiency. Through extensive experiments on multiple medical datasets, we assess the feasibility of state-of-the-art distillation techniques, demonstrating that condensed datasets can achieve near-equivalent model performance compared to full datasets. Moreover, we introduce predictive indicators to evaluate distillation effectiveness across various imaging modalities, paving the way for more efficient and secure medical data sharing. In parallel, we propose an automatic image unfolding and stitching framework for esophageal endoscopy to overcome the limitations of narrow-field imaging. Our method integrates depth-based unfolding, feature-matching techniques (LoFTR, SIFT, ORB), and a novel Density-Weighted Homography Optimization (DWHO) algorithm to enhance image alignment and stitching accuracy. By reconstructing high-quality panoramic views of the esophagus, our approach improves diagnostic clarity and clinical decision-making. Together, these contributions address critical bottlenecks in medical imaging by enhancing data efficiency and visualization fidelity, thereby advancing computational approaches for medical diagnostics and collaborative research.