Combination of Activation Functions Within LeNet-5 Framework for Autism Screening

Recent advancements in deep learning have ushered in innovative methodologies for analyzing extensive datasets, particularly in the medical field. Among the various models, Convolutional Neural Networks (CNNs), including the renowned LeNet-5 model, have emerged as leading supervised deep learning models for classifying functional brain images. In the realm of neuroimaging, functional Magnetic Resonance Imaging (fMRI) holds particular significance, especially in the classification of Autism Spectrum Disorder (ASD), a severe neurodevelopmental disorder affecting the infant population. There exists a crucial unmet need for early autism screening, and fMRI, as a non-invasive brain imaging technique, presents a promising avenue for addressing this need. The proposed work explores potential approaches for timely autism intervention, providing insights into recent advances in ASD classification using a 3D CNN architecture for multimodal fMRI data. To enhance the robustness of the LeNet-5 model, a combination of the best-performing activation functions is employed along with ensemble learning. Experimental results unequivocally demonstrate that the proposed ensemble surpasses the performance of the LeNet-5 model using the benchmark ReLU activation function and other standalone activation functions. To ensure reliable performance comparison, the proposed framework undergoes quantitative testing on all datasets from the ABIDE I database, adopting the LeNet-5 model. The superiority of the proposed system is evident, achieving a notable 92% accuracy with the CMUa dataset.