Deep learning-based synthetic brain MRI for the assessment of regional atrophy patterns in neurodegenerative diseases.

objectivesAssessing regional brain atrophy on 3D-T1w imaging is crucial for evaluating neurodegenerative disorders. However, high-quality volumetric imaging is not always available. Thus, AI-based algorithms were developed to generate "synthetic" 3D-T1w sequences using various clinical sequences as input. This retrospective study aims to investigate whether regional atrophy patterns are preserved in deep learning-based synthetic 3D-T1w sequences from different inputs. MATERIALS AND

methodsThe study included patients with Alzheimer's disease (AD), Frontotemporal dementia (FTD), and healthy controls (HC). Probands were scanned at 3 T, and deep learning-based synthetic 3D-T1w images were generated from various inputs (3D FLAIR, 4 mm axial FLAIR, 4 mm coronal T2) using FreeSurfer-based SynthSR. Real 3D-T1w images served as the reference standard. Brain volumetry was performed using SynthSeg+ in FreeSurfer and the AssemblyNet-AD-FTD pipeline in VolBrain.

resultsGlobal and regional volumes differed significantly between deep learning-based synthetized sequences and the reference standard 3D T1 for all subgroups and inputs (total white matter volume AD p = 0.0002, FTD p < 0.0001, HC p = 0.0116; total gray matter volume for AD, FTD, and HC p < 0.0001), except for hippocampal volumes. This systematic error in overestimating volumes affected automated disease probability prediction in FTD for all inputs (p < 0.0001) and in HC for coronal T2 input (adj. p = 0.0054).

conclusionDeep learning-based synthetic 3D-T1w sequences introduce systematic errors in assessing global and regional brain volumetric measures, leading to overestimated volumes in controls and patients. Resulting synthetic images should be used cautiously, especially for volumetric analyses. KEY POINTS: Question It remains unclear whether deep learning-based synthetic 3D-T1w images from various inputs preserve regional atrophy patterns sufficiently to serve as input for automated volumetry. Findings Deep learning-based synthetic T1w images overestimate regional and global brain volumes in neurodegenerative diseases and controls, increasing with lower quality inputs. Clinical relevance Deep learning-based synthetic images should only be used with caution for volumetric evaluation of brain MRI scans. If possible, 3D scans should be used as input.