Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that affects behavior, speech and social interaction. Early diagnosis is crucial for effective intervention. This study uses MRI DICOM pictures to classify people with autism and those without using a decentralized deep learning architecture that merges Convolutional Neural Networks (CNNs) with Gossip Learning. Unlike centralized models, each node trains locally and shares only model parameters with random peers, preserving privacy. MRI data is preprocessed through resizing, normalization, and reshaping for CNN input. Nodes train individually and periodically average weights to enable global learning. Evaluations using accuracy, precision, recall, F1-score, and confusion matrices show that Gossip Learning achieves performance comparable to centralized methods, maintaining high classification accuracy. This decentralized approach ensures privacy while providing a scalable solution for ASD detection and medical analysis.