A Multi-Modal Content-Based Medical Image Retrieval Framework Using Marine Predators Optimization and Modality-Aware Attention-Guided Disentangled Contrastive Representation Learning

Medical image retrieval plays an important role in computer-aided diagnosis by enabling access to clinically relevant images from large repositories to support diagnostic decision-making. Despite significant advances, existing Content-Based Medical Image Retrieval (CBMIR) systems often suffer from inadequate modality discrimination, insufficient attention to diagnostically important regions, poor generalization across heterogeneous imaging modalities, and suboptimal feature representations. To address these limitations, this study develops a multi-modal Content-Based Medical Image Retrieval (CBMIR) framework for accurate retrieval of Electroencephalography (EEG), Magnetic Resonance Imaging (MRI), retinal, and X-ray images. The framework incorporates preprocessing and data augmentation, followed by hierarchical feature extraction using Residual Networks with External Transformations (ResNeXt). Modality-Aware Encoding is employed to preserve modality-specific characteristics, while the Convolutional Block Attention Module (CBAM) enhances feature refinement by emphasizing informative regions. In addition, Disentangled Representation Learning, Sparse Feature Learning, and Supervised Contrastive Embedding Learning are integrated to generate compact and highly discriminative feature representations. The Marine Predators Algorithm (MPA) is further utilized to optimize critical hyperparameters and improve retrieval effectiveness. During retrieval, Cosine Similarity is applied to rank database images and identify the most relevant matches. Experimental results demonstrate strong performance, achieving Precision@1 (P@1) of 0.985, Precision@10 (P@10) of 0.958, Normalized Discounted Cumulative Gain@10 (NDCG@10) of 0.964, and Mean Average Precision (mAP) of 0.962, validating the effectiveness of the proposed CBMIR framework for multi-modal medical image retrieval.