Hybrid CNN-RNN Models for Multimodal Analysis of Autism Spectrum Disorder Neuroimaging

ASD is a developmental disorder in the neurological category and has numerous and diverse symptoms. Diagnosis is still problematic because symptoms of parkinsonism are diverse, and the dimensionality of the neuroimaging data is high. This paper presents a combination of CNN and RNN for classifying the ASD by analyzing multimodal neuroimaging data such as fMRI data, MRI data, and DTI data. The CNN component analyzes spatial characteristics of the neuroimaging data, and the RNN learns temporal patterns in fMRI series. Enhancements include a novel preprocessing step called 3D Brain Region Alignment (3D-BRA) as well as the 3D version of Adaptive Temporal, minimizing dimensionality through an innovative feature selection technique employing EN-CC. Predictive analysis with the hybrid model utilizes standard benchmark datasets such as ABIDE I, ABIDE II, HCP, and SFARI and records classification performances over 95 percentage. To enhance the interpretability of the model, the Grad-CAM algorithm is used, which marks specific regions of the brain important in the context of ASD. The proposed framework seems to possess considerable feasibility for enhancing the diagnosis of ASD and improving the understanding of its neural substrates, making the developed method highly advantageous for applications in neuroinformatics and clinical utility.