A machine-learning model to harmonize brain volumetric data for quantitative neuro-radiological assessment of Alzheimer's disease

BackgroundStructural MRI plays a pivotal role in the radiological workup for assessing neurodegeneration. Scanner-related differences hinder quantitative neuroradiological assessment of Alzheimers disease (QNAD). This study aims to train a machine-learning model to harmonize brain volumetric data of patients not encountered during model training. MethodNeuroharmony is a recently developed method that uses image quality metrics (IQM) as predictors to remove scanner-related effects in brain-volumetric data using random forest regression. To account for the interactions between AD-pathology and IQM during harmonization, we developed a multi-class extension of Neuroharmony. We performed cross-validation experiments to benchmark performance against existing approaches using data from 20,864 participants comprising cognitively unimpaired (CU) and impaired (CI) individuals, spanning 11 cohorts and 43 scanners. Evaluation metrics assessed ability to remove scanner-related variations in brain volumes (biomarker concordance), while retaining the ability to delineate different diagnostic groups (preserving disease-related signal). ResultsFor each strategy, biomarker concordances between scanners were significantly better (p < 10-6) compared to pre-harmonized data. The proposed multi-class model achieved significantly higher concordance than the Neuroharmony model trained on CU individuals (CI: p < 10-6, CU: p = 0.02) and preserved disease-related signal better than the Neuroharmony model trained on all individuals without our proposed extension ({Delta}AUC= -0.09). The biomarker concordance was better in scanners seen during training (concordance > 97%) than unseen (concordance < 79%), independent of cognitive status. ConclusionIn a large-scale multi-center dataset, our proposed multi-class Neuroharmony model outperformed other strategies available for harmonizing brain-volumetric data in a clinical setting. This paves the way for enabling QNAD in the future.