A hybrid EEG MRI ResNet50 fusion network with cross modal attention for early Parkinson’s disease detection using REMCAT

Parkinson Disease is a progressive nervous system disorder that causes brain cells that make dopamine to lose its ability to send signals, which causes tremors, stiffness, and slowness. Symptoms start slowly, usually on one side of the body, and get worse over time. It can also have an effect on sleep, mental health, and other things. There is no cure, but medications and therapy can help control the symptoms. In this study, REMCAT is proposed for early Parkinson Disease detection. This model integrates functional EEG spectrograms and structural MRI features through a dual branch deep learning framework. The EEG branch engages a lightweight 2D CNN with three convolutional and pooling layers, followed by Global Average Pooling and a Dense 128 layer to extract compact temporal spectral embeddings. The MRI branch has already trained a ResNet50 backbone. A cross modal attention module uses the fused EEG and MRI characteristics, enabling interaction between the two modalities and the channel gating improves the fused representation, which is then delivered to fully linked layers with SoftMax activation for categorization. The Adam optimizer, categorical cross entropy loss, and callbacks are all utilized to train the model so that it is more stable and works better. The proposed model framework has an accuracy of 98.3%, a precision of 94.8%, a recall of 98%, and a specificity of 98.4%. It also has a Cohen Kappa score of 97% and an AUC ROC of 99% based on augmented, non-independent samples external validation on independent datasets was not performed, showing the classes agree with each other quite well. Subject wise 5 fold Group K Fold cross validation that assures all the subjects are present in either training or testing folder without overlapping giving realistic generalization performances. The hybrid fusion technique is more stable and easier to understand than single modality models. Grad CAM images provide additional confirmation that the model accurately identifies disease relevant regions in both EEG and MRI modalities. The framework constitutes an interpretable, and computationally efficient deep learning model for early Parkinson Disease detection. Through the fusion of EEG and MRI modalities using cross modal attention and transfer learning, it achieves high diagnostic precision while maintaining low computational demand, enabling transparent and real time clinical decision support.