Content-based medical image retrieval enables searching large medical image databases to find visually similar reference cases, aiding tasks like computer-aided diagnosis, treatment planning, and medical research by utilizing prior data. Convolutional neural networks (CNNs) have shown promising results for medical image analysis, but challenging multi-class problems with limited data require enhancements. This work proposes a novel attention enhanced multi-stage residual networks architecture for extracting more complex visual features from medical images to improve image retrieval performance. The key objectives include designing stage-wise residual architectures to learn hierarchically more complex features optimized for similarity matching and integrating domain knowledge to focus learning on diagnostically relevant regions. Public datasets like chest radiography images (X-Rays) are used for feature extraction, evaluation, and image retrieval. The learned feature representations are evaluated using dimensionality reduction and visualization to assess their quality and diagnostic relevance. The extracted features are then integrated into the image retrieval system, enabling the retrieval of similar reference cases. Performance is assessed using standard retrieval metrics.