A Machine Learning Model to Harmonize Volumetric Brain MRI Data for Quantitative Neuroradiological Assessment of Alzheimer Disease.

Abstract

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To extend a previously developed machine learning algorithm for harmonizing brain volumetric data of individuals undergoing neuroradiological assessment of Alzheimer disease not encountered during model training. Materials and Methods Neuroharmony 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 Alzheimer disease pathology and IQM during harmonization, the authors developed a multiclass extension of Neuroharmony for individuals with and without cognitive impairment. Cross-validation experiments were performed to benchmark performance against other available strategies using data from 20,864 participants with and without cognitive impairment, spanning 11 prospective and retrospective cohorts and 43 scanners. Evaluation metrics assessed the ability to remove scanner-related variations in brain volumes (marker concordance between scanner pairs), while retaining the ability to delineate different diagnostic groups (preserving disease-related signal). Results For each strategy, marker concordances between scanners were significantly better (P < .001) compared with preharmonized data. The proposed multiclass model achieved significantly higher concordance (0.75 ± 0.09) than the Neuroharmony model trained on individuals without cognitive impairment (0.70 ± 0.11) and preserved disease-related signal (∆AUC =-0.006 ± 0.027) better than the Neuroharmony model trained on individuals with and without cognitive impairment that did not use our proposed extension (∆AUC =-0.091 ± 0.036). The marker concordance was better in scanners seen during training (concordance > 0.97) than unseen (concordance < 0.79), independently of cognitive status. Conclusion In a large-scale multicenter dataset, our proposed multiclass Neuroharmony model outperformed other available strategies for harmonizing brain volumetric data from unseen scanners in a clinical setting. Published under a CC BY 4.0 license.