<p>Efficient medical image retrieval significantly aids in diagnostic and treatment processes by reducing the time and expertise required for accurate assessments. In this study, a novel two-phase method is proposed to enhance semantic distance in content-based image retrieval tasks using a combination of Siamese neural networks and clustering. The first phase employs K-means clustering to group images based on low-level feature similarities, effectively enhancing the distribution of structurally similar images within clusters. In the second phase, a Siamese neural network refines the similarity measurement, replacing traditional distance metrics with a learned metric that captures high-level semantic features. The proposed method is evaluated on the HAM10000, Lung diseases, and Chest X-ray datasets using three key metrics: Precision@K, mean Average Precision at K (mAP@K), and F1@K. Results indicate that our approach consistently outperforms traditional distance metrics across all evaluation metrics, demonstrating its effectiveness in providing more accurate and reliable image retrieval for medical applications.</p>