Self-Evolving Neural Imaging Models For Automated Disease Localization In Multimodal Scans

Early and precise localization of abnormal regions in multimodal medical imaging—such as MRI, CT, PET, and ultrasound—is essential for diagnosing complex and rare cardiovascular diseases. Traditional deep learning models often require large annotated datasets and frequent manual tuning, which limits scalability in clinical environments where annotated samples for rare cardiac abnormalities are scarce. To address this challenge, this study proposes a Self-Evolving Neural Imaging Model capable of autonomously adapting its architecture, parameters, and feature-extraction strategies for automated disease localization across heterogeneous imaging modalities. The proposed model integrates evolutionary neural optimization, multimodal attention fusion, and adaptive self-reconfiguration to enhance sensitivity to subtle tissue variations associated with rare cardiomyopathies, congenital defects, microvascular alterations, and inflammatory cardiovascular disorders. A hierarchical cross-modal encoder extracts spectral, structural, and intensity-based biomarkers, while an evolutionary controller iteratively improves model accuracy by exploring optimized architectural pathways without human intervention. Experimental evaluation on multimodal cardiovascular imaging datasets demonstrates that the proposed self-evolving system achieves 15–22% improvement in localization accuracy, significant reduction in false positives, and enhanced robustness under noise, motion artifacts, and modality variability. The model requires minimal annotation and displays strong generalization to unseen disease patterns, making it suitable for clinical deployment in early detection and precision diagnostics for rare cardiovascular diseases. This framework contributes to the advancement of intelligent, autonomous imaging systems aligned with JRCD’s mission to improve outcomes through early cardiovascular disease understanding and detection.