NEURO-EVOLUTIONARY ATTENTION NETWORKS FOR DISEASE LOCALIZATION IN MULTIMODAL MEDICAL IMAGES

Abstract

Accurate disease localization in multimodal medical imaging plays a crucial role in computer-aided diagnosis and clinical decision-making. Conventional convolutional neural networks often suffer from limited feature adaptability and reduced robustness when dealing with heterogeneous imaging modalities such as Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Positron Emission Tomography (PET), and Ultrasound scans. To address these limitations, this paper proposes a Neuro-Evolutionary Attention Network (NEAN) framework for automated disease localization in multimodal medical images. The proposed architecture integrates multi-scale convolutional feature extraction with attention mechanisms and an evolutionary optimization strategy to adaptively refine network parameters and enhance lesion localization accuracy. Cross-modal feature fusion is employed to exploit complementary information from different imaging modalities, while a hybrid Genetic Algorithm–Particle Swarm Optimization (GA-PSO) approach is utilized to optimize attention weights and network hyperparameters. Experimental evaluation conducted on multimodal imaging datasets demonstrates that the proposed model achieves a localization accuracy of 98.34%, precision of 97.86%, recall of 98.12%, F1-score of 97.99%, and an AUC of 0.992. Compared with conventional CNN, ResNet50, DenseNet121, and Vision Transformer models, the proposed framework improves localization performance by approximately 3.8–7.4% while reducing computational complexity by 18.6%. The results indicate that the Neuro-Evolutionary Attention Network provides a reliable and efficient solution for intelligent disease localization in multimodal medical imaging applications.