Comparing two artificial intelligence software packages for normative brain volumetry in memory clinic imaging.
Lara A M Zaki1, Meike W Vernooij2,3, Marion Smits1
1Department of Radiology and Nuclear Medicine, Erasmus MC, University Medical Center Rotterdam, PO Box 2040, Rotterdam, 3000 CA, the Netherlands.
Neuroradiology
|January 15, 2022
Summary
Two artificial intelligence software packages for brain MRI volumetry showed moderate diagnostic agreement, impacting dementia diagnosis. Clinics must evaluate AI tools internally, as they are not interchangeable for clinical interpretation.
Area of Science:
- Neurology
- Radiology
- Artificial Intelligence
Background:
- Normative brain volumetry using artificial intelligence (AI) aids in dementia diagnosis.
- Different AI software packages may yield varying results, potentially affecting clinical decisions.
Purpose of the Study:
- To compare two AI software packages for normative brain volumetry.
- To assess the impact of these packages on dementia diagnostics in a clinical setting.
Main Methods:
- Retrospective analysis of 60 patients (Alzheimer's disease, frontotemporal dementia, mild cognitive impairment) and 20 controls.
- Two AI software packages processed MRI scans, generating quantitative reports.
- Neuroradiologists provided diagnoses based on AI output, assessing agreement, accuracy, sensitivity, specificity, and confidence.
Main Results:
- Moderate diagnostic agreement between AI packages for normal/abnormal volumetry (K=0.41-0.43) and specific diagnoses (K=0.36-0.38).
- High inter-observer agreement within each package (K=0.73-0.82).
- No significant differences in diagnostic accuracy, sensitivity, or specificity between packages; differences noted in diagnostic confidence for one rater and in whole brain intracranial volume output (10.73%).
Conclusions:
- AI software packages for brain MRI volumetry can yield distinct clinical interpretations.
- Different AI packages are not interchangeable and require internal validation before clinical adoption.
