EVF-FaNOA: An optimized EFFResNet-ViT deep learning framework for autism spectrum disorder detection using magnetic resonance imaging

Autism Spectrum Disorder (ASD) is associated with neural activity that leads to irregular patterns of early brain development. Although a wide range of conventional approaches exists, these methods rely on observation and behavioural assessment, which often causes ambiguous results. Thereby, EfficientNet-B0, ResNet-50 and Vision Transformer with the Falco-Ninja Optimization Algorithm (EVF-FaNOA) are introduced. Initially, Magnetic Resonance Imaging (MRI) images are obtained from the dataset, and these images undergo pre-processing using a Denoising Autoencoder (DAE). The pre-processed images are then subjected to functional connectivity analysis, where key brain regions are identified using the proposed Falco-Ninja Optimization Algorithm (FaNOA), which is a hybrid of Ninja Optimization (NiOA) and Falco Peregrinus Optimization (FPO). The current output from this respective phase is considered the first output. In parallel, the MRI image undergoes feature extraction and the resulting features, combined with the earlier output, are fed into the ASD detection stage. Here, a hybrid model called EVF-FaNOA, which integrates EfficientNet-B0, ResNet-50 and Vision Transformer (EFFResNet-ViT) with FaNOA, is employed. The EVF-FaNOA is enhanced by incorporating sigmoid cross-entropy and neighbour-based loss functions. The EVF-FaNOA attained a Youden Index (Y-Index) of 92.799 %, accuracy of 95.602 %, and Classification Success Index (CSI) of 93.983 % using image size 512×512.