Deep Learning Approach for ASD Detection with GAN-Driven MRI Data Augmentation

Abstract

Autism spectrum disorder (ASD) is a highly heterogeneous neurodevelopment disorder that is primarily characterized by impaired communication, behaviours and social interaction. These will remain strong commonsense that can not only diagnose ASD accurately but provide as swiftly, especially, when trying to work with small-labelled datasets in the presence of neuro-imaging data such as MRI photos, etc. Here, we present a hybrid deep learning framework combining CNNs and Swin attention maps to discriminate MRI data from different groups of children with a diagnosis of autism spectrum disorder (ASD) based on the complementarity of the features they extract. Furthermore, we introduce a GAN-based synthetic data augmentation method to tackle the data scarcity issue. Here, we’re operating on the slices already prepared for them from each MRI scan, again crop the middle slice from each, and then wrap-in the needed resize so that it could be input to a model. This results in a bigger/ broader training set and conditioning of the GAN generator to produce synthetic images that closely resemble typical ASD neuroimaging data. Similarly, the combined CNN model provides a short attention and large feature extraction method depends on the depth-wise convolution and a comprehensive "lossless" classification approach depends on the layered parts. Trained and validated on MRI slices of ASD and non-ASD subjects, synthetic data generated a superlative model. Hence, we have metrics e.g. accuracy, precision, recall, AUC-ROC to check performance of the model and our validation accuracy score is 75%+. Experiments results show improvement in performance on ASD classification by data augmentation using synthetic data and attention mechanisms. This does, however, appears to be the most unfortunate prevalent approach to the problem, as the annotation of images has too high resource requirements that make this often pointless like most deep learning, a lot of machines learning just end up wasted and a lot to be cleaned up, becoming clear from all the papers we review that undergo the problem with a real-life mindset. In the future, we could explore 3D CNN research with similar methods and examine stronger attention mechanisms, which may yield an incremental improvement in identification performance for ASD. The results from our finding can help an adaptation on machine learn based to apply for neurodevelopment disorder disease’s diagnosis purpose potentially and raise the feasibility and access to the diagnostic tools.