Human visual nigrosome analysis improves AI-based diagnoses in neurodegenerative disease
Author Block: E. Sayilir1, E. Piot1, C. Bonardel1, F. Renard2, S. Grand1, A. Attye2, A. Krainik1; 1Grenoble/FR, 2La Tronche/FR
Purpose: This study aimed to evaluate the diagnostic performance of BrainGML a manifold-learning AI software (Geodaisics.com) in neurodegenerative diseases combined with the visual analysis of nigrosome imaging. Neurodegenerative diseases are associated with regional cerebral atrophy patterns, such as temporohippocampal and parietal atrophy in Alzheimer's disease (AD), and frontotemporal atrophy in frontotemporal dementia (FTD). Parkinson's (PD) and dementia with Lewy bodies (DLB) have subtle structural abnormalities such as insular atrophy and nigrosome loss. Artificial intelligence software shows potential for accurate diagnosis of atrophy patterns, while nigrosome analysis is still unavailable.
Methods or Background: A retrospective study was conducted on patients with AD, FTD, PD, and DLB. BrainGML analysed
cerebral atrophy using 3DT1 images, assigning the highest probability for Normal, AD, FTD, or PD. Nigrosome visual analysis was performed on susceptibility-weighted images by 4 radiologists who determined whether nigrosomes were normal or abnormal. The primary outcome was the accuracy of the radiological diagnosis, defined as 'Right', 'Wrong', or 'Undefined' (when nigrosome imaging was normal in PD and DLB, or when nigrosome imaging was abnormal in AD or FTD).
Results or Findings: The cohort included 79 patients (29 AD, 11 FTD, 26 PD, 13 DLB). Nigrosomes were normal in
AD (100%) and FTD (100%), and were abnormal in PD (92%) and DLB (62%). BrainGML provided 59% Right diagnoses on trained diagnoses (all but DLB), and 49% on all patients including DLB. Combining nigrosome analysis with BrainGML decreased the ratio of 'Wrong' diagnoses from 51% to 15%, which was replaced by the increase of 'Undefined' diagnoses from 0 to 36%.
Conclusion: In conclusion, adding nigrosome visual analysis to BrainGML highest diagnosis probability turned most 'wrong' diagnoses into 'undefined' diagnoses.
Limitations: The main limitations are small population samples, unknown DLB diagnosis by BrainGML, and no pathological diagnoses.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the French South-East Ethics committee.