Hybrid Deep Learning Approaches for Classifying Autism from Brain MRI
This work addresses improving objective diagnosis of autism spectrum disorder for clinicians and researchers, but it is incremental as it combines existing methods on a known dataset.
The researchers tackled classifying autism from brain MRI by testing a 3D CNN and a hybrid CNN+SVM approach on the ABIDE I dataset, finding that the hybrid model achieved higher accuracy (0.76) and AUC (0.80) compared to the baseline CNN (accuracy 0.66, AUC 0.70).
Autism spectrum disorder (ASD) is most often diagnosed using behavioral evaluations, which can vary between clinicians. Brain imaging, combined with machine learning, may help identify more objective patterns linked to ASD. This project used magnetic resonance imaging (MRI) data from the publicly available ABIDE I dataset (n = 1,112) to test two approaches for classifying ASD and control participants. The first was a 3D convolutional neural network (CNN) trained end-to-end. The second was a hybrid approach that used the CNN as a feature extractor and then applied a support vector machine (SVM) classifier. The baseline CNN reached moderate performance (accuracy = 0.66, AUC = 0.70), while the hybrid CNN + SVM achieved higher overall accuracy (0.76) and AUC (0.80). The hybrid model also produced more balanced results between ASD and control groups. Separating feature extraction and classification improved performance and reduced bias between diagnostic groups. These findings suggest that combining deep learning and traditional machine learning methods could enhance the reliability of MRI-based research on ASD.