Automated MRS quality control across diverse acquisitions and neurological diseases using self-supervised transformers

Motivation: Preprocessing pipelines for MRSI are time-consuming and typically require trained experts with domain knowledge. With accelerated acquisitions and improved spatial coverage, automated quality control (QC) becomes increasingly important. Goal(s): To develop a model for automated spectral QC that captures global brain spectra characteristics while being robust across diverse neurological diseases Approach: We employed a contrastive self-supervised learning framework during pretext training phase. A classification model was stacked on pretrained latent space to predict the quality of voxel-wise MRS data. Results: The proposed model showed its strength in modeling complex dependencies over spectral sequences and demonstrated robustness across diverse neurological disease and scanner acquisitions Impact: This approach enables automated quality control for MRS, reducing reliance on manual assessments. It enhances diagnostic accuracy, supports broader clinical adoption, and may reveal complex interdependencies, thus improving our understanding of neurological disease progression and treatment responses