Multiparametric Radiomics-Informed Transformer Framework for Autism Spectrum Disorder Diagnosis

MICCAI 2024 Workshop MSB Submission9 Authors

Keywords: Corpus Callosum, Autism Spectrum Disorder, Radiomics, MRI, Deep Learning

TL;DR: We propose a new approach for CC segmentation from brain MRIs based on radiomic and deep learning techniques.

Abstract:

Autism Spectrum Disorder (ASD) is one of the leading neurodevelopmental disorders in the world and rapidly increasing in prevalence. Existing automated ASD prediction systems have two critical drawbacks. The first one involves making the prediction using only Corpus Callosum (CC) segmentation data, whereas since similar changes are seen in other mental illnesses like schizophrenia and bipolar disorder, the CC segmentation is not enough to predict ASD. The second issue is that single parameteric from neuroimaging data; such as volume, area, or fiber integrity; cannot fully capture the intricate neuropathological changes in the CC associated with ASD development. There is no multiparametric-based radiomics learning model that is based solely on a single ROI.
To address these limitations, we propose a radiomics-informed transformer framework for detecting ASD from a single ROI-based radiomics extracted features. The proposed framework operates through two key mechanisms. First, we have developed an optimized hidden Markov random field algorithm for CC segmentation that addresses resource constraints by focusing exclusively on the localized region of the CC. Second, we leverage BERT to distinguish radiomic features of healthy and ASD subjects. Furthermore, we ensure complementary information is learned by tokenized radiomics and radiomic features by designing an effective feature de-correlation loss. Combined, our method addresses the limitations of ASD diagnosis, achieving 98.2% DSC for CC segmentation and 96.8% for the ASD classification task.

Submission Number: 9