Leveraging TabTransformer Deep Learning on Conventional MRI Radiomics for Accessible and Interpretable Diagnosis of...

This study aims to assess the diagnostic efficacy of a multi-region radiomics analysis utilizing conventional MRI sequences (T1-weighted imaging [T1WI] and T2-weighted imaging [T2WI]) for autism spectrum disorder (ASD), and to investigate the correlations between radiomics features and the severity of clinical symptoms, thereby exploring potential imaging biomarkers.

This retrospective study included 207 pediatric participants (91 ASD, 116 typically developing controls). Radiomics features were extracted from manually segmented bilateral hippocampus, thalamus, caudate nucleus, and lenticular nucleus on T1WI and T2WI images. Three distinct classifiers (T1WI-only, T2WI-only, T1WI+T2WI combined) were developed using logistic regression (LR), support vector machine (SVM), and a TabTransformer deep learning (DL) model. Diagnostic performance was evaluated via five-fold cross-validation.

The TabTransformer DL model utilizing combined T1WI+T2WI features demonstrated superior performance, achieving an area under the curve of 0.900, accuracy of 0.834, sensitivity of 0.843, and specificity of 0.823. Specific radiomic features, predominantly from the left lentiform nucleus and bilateral caudate nucleus, were significantly correlated with clinical severity scores (ABC, CARS).

Radiomics models leveraging routine MRI sequences demonstrate robust diagnostic utility for ASD. The identified subcortical features, correlating with core symptoms, may serve as viable imaging biomarkers. Future work requires external validation, exploration of automated segmentation, and investigation in larger, multi-center cohorts..