NeuroASD-Net: A Deep Learning-Based Approach for Autism Detection from Structural MRI

Autism spectrum disorder (ASD), a complex neurodevelopmental disorder, is typified by social interaction challenges, communication difficulties, and repetitive activities. Effective intervention requires an early and precise diagnosis. This paper introduces NeuroASD-Net, a novel framework for ASD detection from structural Magnetic Resonance Imaging (MRI). The approach integrates advanced preprocessing techniques such as histogram matching and Contrast Limited Adaptive Histogram Equalization (CLAHE) to standardize input data. The next stage is the segmentation of key brain regions, like cerebrospinal fluid, white matter, and gray matter, using k-Means clustering. This segmentation step isolates critical regions for subsequent analysis. Following segmentation, a Two-Level Feature Extraction (TLFE) model is applied. In the first level, spatial and textural patterns are extracted using Local Binary Patterns (LBP), Gabor filters, and the Gray Level Co-occurrence Matrix (GLCM), capturing essential morphological characteristics of the brain. Feature selection is then optimized using the Self-Adaptive Black-Winged Kite Algorithm (SA-BKA). In the second level, high-level features are extracted using EffiDSR-Net, in which the EfficientNet model is enhanced with Dual Scale Residual (DSR) Blocks and a Convolutional Block Attention Module (CBAM). The features from both levels are fused to form a comprehensive feature set. Finally, classification is performed using a Softmax layer, achieving precise ASD detection. The proposed framework demonstrates enhanced diagnostic accuracy, achieving an Accuracy of 99.28%, Sensitivity of 99.31%, Specificity of 99.24%, F1-score of 99.31%, and an Area Under the ROC Curve (AUC) of 0.995, indicating its effectiveness for clinical ASD detection.