Peer-reviewed article
Multiple classification of childhood and juvenile autism spectrum disorder from ABIDE-II sMRI using BN-GO-CNN.
Yu Ke, Qingliang He, Wei Jiang, Yusi Chen, Yueqian Ke ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This is a technical study about using artificial intelligence to classify autism spectrum disorder from brain MRI scans. It develops and tests a computer algorithm to distinguish between children and adults with autism based on structural brain imaging. The algorithm achieved modest accuracy levels (60-67%) and the authors conclude it is not yet reliable for clinical diagnosis. The work uses data from an existing autism imaging dataset and compares different machine learning approaches.
Key points
- A new CNN algorithm (BN-GO-CNN) was developed to classify autism cases by age and sex using structural MRI brain scans
- Validation accuracies ranged from 60.5% to 67.0% across different classification tasks, remaining above chance but with limited class separability
- The modest performance suggests structural MRI differences related to age and sex in autism are subtle
- The authors explicitly state the framework is 'not clinically reliable' and intended only for future research, not clinical diagnosis
- The study used ABIDE-II dataset and compared performance against established deep learning models (VGG16, DenseNet201, MobileNetV2, EfficientNet-B0)
Limitations
This is a methodological technical study focused on algorithm development rather than evaluating an intervention to improve child well-being. It does not assess health outcomes, service delivery, or impacts on children's lives—only the technical performance of machine learning models on imaging data.
Abstract
OBJECTIVES: Autism spectrum disorder (ASD) shows heterogeneous neurodevelopmental patterns across childhood, juvenile development, and adulthood, so binary MRI classification may not fully capture age-and sex-related variation. This study proposes a proof-of-concept framework for multiple classification using ABIDE-II T1-weighted structural MRI (sMRI). METHODS: After image-quality screening, sMRI slices were processed using Canny edge detection, brain-region cropping, 224 × 224 resizing, and augmentation. Images were organized into sex-based four-class, age-based four-class, and combined age-and-sex eight-class datasets, separating child, young, juvenile, adult, male, and female groups. A Batch-Normalized Grid-Optimized CNN (BN-GO-CNN) was developed to learn structural representations directly from preprocessed sMRI, with grid search used to select key CNN hyperparameters. The model was compared with VGG16, DenseNet201, MobileNetV2, and EfficientNet-B0 using identical image partitions. RESULTS: BN-GO-CNN achieved validation accuracies of 67.0 %, 64.0 %, and 60.5 % for sex-based, age-based, and combined age-and-sex classification, respectively. These outcomes remained above chance but showed limited class separability. CONCLUSIONS: The modest performance suggests that label-related sMRI differences are subtle, supporting this framework as a preliminary test of CNN-assisted sMRI classification for future research use rather than a clinically reliable ASD diagnostic system.