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

Area of Science:

• Neuroimaging
• Radiomics
• Biomarkers

Background:

• Autism spectrum disorder (ASD) diagnosis relies on behavioral assessments.
• Identifying objective imaging biomarkers for ASD is crucial for early detection and intervention.

Purpose of the Study:

• To evaluate the diagnostic accuracy of multi-region radiomics using conventional MRI (T1WI, T2WI) for ASD.
• To explore correlations between radiomics features and ASD clinical symptom severity.
• To identify potential neuroimaging biomarkers for ASD.

Main Methods:

• Retrospective analysis of 207 pediatric participants (91 ASD, 116 controls).
• Radiomics features extracted from segmented subcortical regions (hippocampus, thalamus, caudate, lenticular nucleus) on T1WI and T2WI.
• Classifiers developed using logistic regression, SVM, and TabTransformer deep learning (DL) with T1WI-only, T2WI-only, and combined T1WI+T2WI features.
• Performance assessed via five-fold cross-validation.

Main Results:

• The TabTransformer DL model with combined T1WI+T2WI features achieved high diagnostic performance (AUC=0.900, accuracy=0.834, sensitivity=0.843, specificity=0.823).
• Significant correlations were found between specific radiomic features (left lentiform nucleus, bilateral caudate nucleus) and clinical severity scores (ABC, CARS).

Conclusions:

• Radiomics models using standard MRI sequences offer effective diagnostic utility for ASD.
• Identified subcortical radiomic features correlate with core ASD symptoms, suggesting potential as imaging biomarkers.
• Further validation in larger, multi-center studies and exploration of automated segmentation are warranted.