Graph Feature Embeddings for Patient Re-Identification from Chest X-Ray Images

Patient re-identification in medical imaging facilitates longitudinal studies, monitors treatment, and ensures patient privacy. Accurate patient re-identification enables clinicians to track patient progress, compare new imaging results with historical data, and ensure that the correct treatment plans are followed without compromising patient confidentiality. However, identifying similar patients presents significant challenges when dealing with low-quality images, like chest X-ray images, especially when the presence of medical equipment obscures key anatomical features. This paper introduces a Graph Matching Network-based approach for patient re-identification using chest X-ray data. By representing such images as graphs, where nodes correspond to key anatomical landmarks and edges represent spatial relationships, the Graph Matching Network can effectively model the complex dependencies within the images. In addition, we integrate superpixel as a representative feature extraction approach in a robust strategy to describe the images as a graph model. Our method is evaluated on a large-scale dataset of chest X-ray images, demonstrating its superior performance compared to other methods. Experimental results show that our approach improves the precision of patient matching by integrating a novel loss function based on the cosine distance of the graph embedding representation, enhancing its robustness against common challenges such as variations in image quality, patient posture, and imaging equipment.