Medical Image Similarity Matching Using Deep Learning

Content-Based Image Retrieval (CBIR) is an emerging tool in medical imaging, enabling clinicians to retrieve visually similar cases for reference, diagnosis, and decision support. However, conventional CBIR systems often struggle with domain variability, image noise, and the semantic gap between low-level visual features and clinical relevance. In this study, we propose a hybrid CBIR framework that integrates adaptive image preprocessing with DINOv2—a self-supervised Vision Transformer pretrained on large-scale natural image datasets—for high-quality feature extraction. To enhance anatomical structure visibility and reduce image noise, input images are filtered using a combination of Gaussian and Laplacian techniques before being embedded by DINOv2. Retrieved images are ranked using cosine similarity via Facebook AI Similarity Search (FAISS). Experiments on the ChestMNIST dataset demonstrate that DINOv2 achieves superior retrieval performance (mAP: 0.25, P@5: 0.60) compared to MedViT and Swin Transformer, despite no task-specific fine-tuning. These results highlight the potential of general-purpose self-supervised models like DINOv2 in medical CBIR, offering high transferability and robustness to varying image quality without requiring extensive domain adaptation.