A Multi-Atlas Enhanced Transformer Framework for Autism Spectrum Disorder Classification
Autism Spectrum Disorder (ASD) is a common psychiatric condition characterized by atypical cognitive, emotional and social patterns. With millions of people affected worldwide, early diagnosis is essential for positive outcomes. However, current diagnostic methods, which rely on behavioral observations, are time-consuming, subjective, and require trained clinicians. In addition, common assessments of ASD lack objectivity and transparency.
Recent advances in magnetic resonance imaging (MRI), particularly functional MRI (fMRI), are providing valuable insights into the pathophysiology of brain disorders. Functional connectivity (FC) analysis, which examines the statistical dependencies between different brain regions, has been instrumental in understanding the network-level abnormalities associated with various disorders. In the context of ASD classification, machine learning approaches applied to resting-state fMRI data have shown promise. However, traditional methods face limitations, especially when dealing with heterogeneous clinical data.
Therefore, we have developed METAFormer, a Multi-Atlas Enhanced Transformer framework designed to improve ASD classification using resting-state fMRI data from the ABIDE I dataset. METAFormer utilizes connectivity matrices from three different brain atlases: AAL, CC200, and DOS160. Notably, our framework incorporates self-supervised pre-training, a process that reconstructs masked values without the need for additional or separate training data.
Our tests with stratified cross-validation on the ABIDE-I dataset show that METAFormer outperforms existing methods. It reaches an average accuracy of 83.7% and an AUC score of 0.832. This is important because it deals with the downsides of usual methods and provides a quick, affordable, and unbiased way to diagnose ASD. We use the ABIDE I dataset, a comprehensive dataset of structural and resting-state fMRI data from individuals with ASD and typical controls. The preprocessing pipeline includes motion correction, intensity normalization, and functional-anatomical registration provided by the Preprocessed Connectomes Project (PCP). Functional connectivity is computed using three different brain atlases: AAL, CC200, and DOS160. The resulting feature vectors are used as input to our models. METAFormer's architecture is based on a multi-atlas transformer, which contains three separate transformers corresponding to different atlases. The self-supervised pre-training task involves the imputation of missing elements in connectivity matrices and is crucial for improving the predictive capabilities of the model for downstream classification.
Our results highlight the effectiveness of METAFormer, not only in outperforming existing classifiers, but also in demonstrating the significant impact of self-supervised pretraining. The model's superior performance, coupled with its ability to address the limitations of traditional methods, underscores its potential to advance objective and efficient ASD diagnosis using advanced neuroimaging and machine learning techniques.