Die Studie untersuchte, ob die Verwendung einer KI-basierten Software die Vorhersage, dass eine Person mit leichten kognitiven Beeinträchtigungen (MCI) Alzheimer entwickeln wird, verbessert. Dafür wurden MRT-Bilder von Radiologen zum einen ohne und zum anderen mit zusätzlichen Informationen der Software ausgewertet. Durch Verwendung des von der Software erstellten Berichts, konnte die Genauigkeit der Vorhersage auf die Entwicklung von Alzheimer verbessert werden.
ECR 2023 EPOS C-20580
V. Richter, J. Steiglechner, U. Ernemann, B. Bender, T. Lindig
Detection of regional brain volume decrease in the diagnostic work up of patients with suspected dementia can be challenging, especially in the early phase without clear clinical symptoms. Automatic quantification and comparison with a reference set of healthy volunteers could be a tool to help neuroradiologist to differentiate normal variance and normal aging from early pathologic changes. Purpose of the presented study was to determine whether detection of brain atrophy and prediction of future conversion of mild cognitive impairment (MCI) to Alzheimer dementia (AD) is improved by utilizing an AI-based software tool for regional volumetry and the effect of rater expertise.
Methods and materials
129 3D-T1-weighted baseline MRI datasets from the ADNI database were presented to 4 radiologists (experience in neuroradiology ranging from 3 months to 15 years) who were blinded to all clinical data, first reviewing the MRI only, then 8 weeks later with adding the regional volumetry report generated by the software AIRAscore (v2.0.1, AIRAmed GmbH, Tübingen) in a randomized order.The dataset contained patients without conversion to AD and patients with a conversion to AD at a later time point. At time of acquisition of the presented MR images all patients were diagnosed MCI. The readers determined whether there is any atrophy present and whether there is atrophy suggestive of AD. Interrater reliability with Fleiss Kappa and reading accuracy were calculated.
A t-test found no significant age difference between subjects with and without future conversion in the examined cohort (mean age 74.5 years, 51:78 female:male, 60 subjects with conversion to AD). In the subgroup of subjects with conversion, mean time between imaging and conversion was 1.98 years (range 0.5-4.19 years). Interrater reliability was significantly improved by adding the AI report both in determining any atrophy (0.56 vs 0.38) and atrophy suggestive of AD (0.77 vs 0.49). A significant improvement in readers` accuracy from 55% to 67% was found by adding the regional volumetry reports (p=.007) with a sensitivity of 57% and specificity of 75%. If only subjects with a conversion within 500 days were considered in the subgroup of converters, the effect was more pronounced with an improvement in readers accuracy from 62% to 77% (p<.001), with a sensitivity of 74% and specificity of 83%. This improvement was not dependent on the experience level of the readers. Mean total reading time in minutes with the report decreased from 115 minutes to 101 minutes but the effect did not reach significance (p=.18).
The addition of an AI-based regional volumetry tool improves accuracy of atrophy detection and prediction of future conversion to AD in the review of MRI datasets of MCI patients. This is in line with the results from other tools (e.g. MorphoBox) that also showed an increase of the inter-rater agreement (kappa between 0.36-0.56 vs 0.55-0.65) with the use of a quantitative report . Also sensitivity of around 75%-83% is in line with comparable evaluation of other groups with different tools . As accuracy was clearly higher for patients with a conversion within the next 16 months, single time point analysis seem insufficient to detect a potential conversion to AD in an early time frame. Atrophy rates might be a potential additional feature that could help here and are currently evaluated.