Clinical neuroradiology · 2026
Swarm Learning for the Detection of Middle Cerebral Artery Occlusion On CT Angiography.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Harten Julia K, Hansel Soma, Creuzberg Dominik, Brugnara Gianluca, Kuetting Daniel, Uebachs Mischa, Weller Johannes, Petzold Gabor C, Dorn Franziska, Vollmuth Philipp, Aschenbrenner Anna C, Schultze Joachim L, Radbruch Alexander, Effland Alexander, Lehnen Nils C
Аннотация
PURPOSE: Swarm Learning has been introduced as a decentralized alternative to Central Learning. We aimed to evaluate whether Swarm Learning (SL) is equivalent to Central Learning (CL) for training AI models to detect middle cerebral artery (MCA) occlusions on CT angiography (CTA).
METHODS: This retrospective study included CTAs of patients suspected of ischemic stroke and negative controls from two institutions and three scanners between October 2010 and March 2024. After manual segmentation of occluded vessels by two raters, two AI algorithms (MedNeXt‑S, Swin UNETR) were trained using CL and SL. Sensitivity, specificity, positive/negative predictive values (PPV, NPV), accuracy, and AUC were calculated. Significance testing was performed using the exact McNemar test and comparisons of relative predictive values. For external validation, CTAs from a fourth scanner from a third institution were used.
RESULTS: 470 CTAs of 470 patients (mean age, 72.6 years ±14.3; 247 women) were included, with 356 (75.7%) MCA-occlusions. Training used 375-377 CTAs (79.8-80.2%), with 93-95 (19.8-20.2%) for internal validation. For internal validation, averaged accuracy ranged from 87.23-92.54%, sensitivity from 91.13-93.56%, specificity from 64.82-95.92%, PPV from 90.27-98.44%, and NPV from 73.39-78.66%. For external validation, 101 CTAs of 101 patients (mean age 70.9 years ±16.8; 56 women) were included, with 27 CTAs (26.7%) showing MCA occlusion. Neither internal nor external validation showed significant differences between SL and CL.
CONCLUSION: AI models trained with SL showed no significant differences compared with CL and may represent a sustainable, and data-safe alternative.
METHODS: This retrospective study included CTAs of patients suspected of ischemic stroke and negative controls from two institutions and three scanners between October 2010 and March 2024. After manual segmentation of occluded vessels by two raters, two AI algorithms (MedNeXt‑S, Swin UNETR) were trained using CL and SL. Sensitivity, specificity, positive/negative predictive values (PPV, NPV), accuracy, and AUC were calculated. Significance testing was performed using the exact McNemar test and comparisons of relative predictive values. For external validation, CTAs from a fourth scanner from a third institution were used.
RESULTS: 470 CTAs of 470 patients (mean age, 72.6 years ±14.3; 247 women) were included, with 356 (75.7%) MCA-occlusions. Training used 375-377 CTAs (79.8-80.2%), with 93-95 (19.8-20.2%) for internal validation. For internal validation, averaged accuracy ranged from 87.23-92.54%, sensitivity from 91.13-93.56%, specificity from 64.82-95.92%, PPV from 90.27-98.44%, and NPV from 73.39-78.66%. For external validation, 101 CTAs of 101 patients (mean age 70.9 years ±16.8; 56 women) were included, with 27 CTAs (26.7%) showing MCA occlusion. Neither internal nor external validation showed significant differences between SL and CL.
CONCLUSION: AI models trained with SL showed no significant differences compared with CL and may represent a sustainable, and data-safe alternative.