PT  - JOURNAL ARTICLE
AU  - Spampinato, Maria Vittoria
AU  - Collins, Heather R.
AU  - Wells, Hannah
AU  - Dennis, William
AU  - Chamberlin, Jordan H.
AU  - Ye, Emily
AU  - Chetta, Justin A.
AU  - Matheus, Maria Gisele
AU  - Stalcup, Seth T.
AU  - Roberts, Donna R.
TI  - Cross-Sectional Validation of an Automated Lesion Segmentation Software in Multiple Sclerosis: Comparison with Radiologist Assessments
AID  - 10.3174/ajnr.A8655
DP  - 2025 Jul 01
TA  - American Journal of Neuroradiology
PG  - 1510--1516
VI  - 46
IP  - 7
4099  - http://www.ajnr.org/content/46/7/1510.short
4100  - http://www.ajnr.org/content/46/7/1510.full
SO  - Am. J. Neuroradiol.2025 Jul 01; 46
AB  - BACKGROUND AND PURPOSE: MRI is widely used to assess disease burden in MS. This study aimed to evaluate the effectiveness of a commercially available k-nearest neighbors (kNN) network software in quantifying white matter lesion (WML) burden in MS. We compared the software’s WML quantification to expert radiologists’ assessments.MATERIALS AND METHODS: We retrospectively reviewed brain MRI examinations of adult patients with MS and of adult patients without MS and with a normal brain MRI referred from the neurology clinic. MR images were processed by using an AI-powered, cloud-based kNN software, which generated a DICOM lesion distribution map and a report of WML count and volume in 4 brain regions (periventricular, deep, juxtacortical, and infratentorial white matter). Two blinded radiologists performed semiquantitative assessments of WM lesion load and lesion segmentation accuracy. Additionally, 4 blinded neuroradiologists independently reviewed the data to determine if MRI findings supported an MS diagnosis. The associations between radiologist-rated WML load and kNN model WML volume and count were evaluated with Spearman rank order correlation coefficient (rho) because these variables were not normally distributed. Results were considered significant when P < .05.RESULTS: The study included 32 patients with MS (35.4 years ±9.1) and 19 patients without MS (33.5 years ±12.1). The kNN software demonstrated 94.1% and 84.3% accuracy in differentiating MS from non-MS subjects based respectively on WML count and WML volume, compared with radiologists’ accuracy of 90.2% to 94.1%. Lesion segmentation was more accurate for the deep WM and infratentorial regions than for the juxtacortical region (both P < .001).CONCLUSIONS: kNN-derived WML volume and WML count provide valuable quantitative metrics of disease burden in MS. AI-powered postprocessing software may enhance the interpretation of brain MRIs in MS patients.3D DIR3D double-inversion recoveryAUCarea under the curveEDSSExpanded Disability Status ScaleFNfalse-negativeFPfalse-positiveICCintraclass correlation coefficientkNNk-nearest neighborsMP2RAGEmagnetization prepared 2 rapid acquisition gradient echoesROCreceiver operating characteristicSPACEsampling perfection with application-optimized contrasts by using a different flip angle evolutionWMHwhite matter hyperintensityWMLwhite matter lesion