European journal of radiology · 2026
Vertebral body segmentation in CT: An open dataset, deep-learning models and comparison to existing models.
СтатьяСравнительное исследованиеНазвание 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.
Hofmann Felix O, Auhage Liv A, Dexl Jakob, Tschaidse Tengis, Pretzsch Elise, Aghamaliyev Ughur, Kunz Wolfgang G, Ingrisch Michael, Renz Bernhard W, Werner Jens, Hoppe Boj F, Ilmer Matthias
Аннотация
PURPOSE: Vertebral bodies are anatomical landmarks for localizing measurements in applications such as body composition analysis. This study aimed to provide open-access vertebral body labels and deep-learning segmentation models, and to compare their performance in identifying the third lumbar vertebra (L3) with existing solutions.
MATERIALS AND METHODS: Thoracic and lumbar vertebral body labels were created for 1460 CT scans from two public datasets. Two residual-encoder nnU-Net models were trained on 1216 cases, and a two-step pipeline was developed. Segmentation performance was tested on 244 cases. In an independent oncological dataset of 300 cases, three readers manually annotated L3 landmarks. Labeling performance was assessed using signed and absolute center localization error (cranio-caudal distance between predicted and annotated L3 center) and center hit rate (predicted center within annotated vertebral body), followed by a comparison with existing models.
RESULTS: The two models and the two-step pipeline achieved Dice scores of 0.962 [95% CI, 0.941 to 0.978], 0.939 [95% CI, 0.907 to 0.965], and 0.953 [95% CI, 0.926 to 0.974] for segmentation of the L3 vertebral body. Center hit rates were similarly high across models. Signed center localization error did not differ significantly, whereas absolute center localization error was lower for VertebralBodiesL and VertebralBodiesStepwise than for the vertebral-body-specific TotalSegmentator pipeline. SPINEPS / VERIDAH results are reported separately because vertebral definitions differed.
CONCLUSION: The proposed models accurately segment thoracic and lumbar vertebral bodies and reliably identify L3. Labels, model weights, and a body composition analysis pipeline are openly available.
MATERIALS AND METHODS: Thoracic and lumbar vertebral body labels were created for 1460 CT scans from two public datasets. Two residual-encoder nnU-Net models were trained on 1216 cases, and a two-step pipeline was developed. Segmentation performance was tested on 244 cases. In an independent oncological dataset of 300 cases, three readers manually annotated L3 landmarks. Labeling performance was assessed using signed and absolute center localization error (cranio-caudal distance between predicted and annotated L3 center) and center hit rate (predicted center within annotated vertebral body), followed by a comparison with existing models.
RESULTS: The two models and the two-step pipeline achieved Dice scores of 0.962 [95% CI, 0.941 to 0.978], 0.939 [95% CI, 0.907 to 0.965], and 0.953 [95% CI, 0.926 to 0.974] for segmentation of the L3 vertebral body. Center hit rates were similarly high across models. Signed center localization error did not differ significantly, whereas absolute center localization error was lower for VertebralBodiesL and VertebralBodiesStepwise than for the vertebral-body-specific TotalSegmentator pipeline. SPINEPS / VERIDAH results are reported separately because vertebral definitions differed.
CONCLUSION: The proposed models accurately segment thoracic and lumbar vertebral bodies and reliably identify L3. Labels, model weights, and a body composition analysis pipeline are openly available.