Journal of applied clinical medical physics · 2026
Transferability of radiomics models between deep learning and conventional CT reconstruction algorithms: A task-based assessment for stratifying acute pancreatitis severity.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
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.
Hu Xiaobao, Liang Xiao, Huang Zhongren, Zhu Yanyan, Guo Mengya, Wu Zhihua, Zhou Fuqing
Аннотация
BACKGROUND: Extensive evidence has demonstrated the superior image quality achieved through deep learning-based reconstruction algorithms. However, given their growing adoption in clinical imaging, the impact of these algorithms on radiomics model development warrants thorough investigation.
PURPOSE: To investigate the transferability of radiomics models for acute pancreatitis (AP) severity stratification across CT images reconstructed with filtered back projection (FBP), adaptive statistical iterative reconstruction-Veo (ASIR-V), and deep learning image reconstruction (DLIR) algorithms.
METHODS: This retrospective study enrolled 79 AP patients who underwent contrast-enhanced CT. Five sets of images were reconstructed with FBP, ASIR-V at 50% blending level (AR50), and DLIR at low (DLRL), medium (DLRM), and high (DLRH) strength levels. A total of 837 radiomic features were extracted from the pancreatic parenchyma volume to generate five corresponding datasets. Radiomics models for stratifying AP severity were built on four machine learning algorithms, trained on three DLIR datasets (source datasets) and tested on AR50 and FBP datasets (target datasets). The source and target datasets were then swapped for reverse evaluation. Model transferability across datasets was assessed by comparing performance differences between source and target datasets, with area under the receiver operating characteristic curve (AUC) as the primary metric.
RESULTS: We identified a fundamental asymmetry in model transferability. Models trained on conventional reconstructions (AR50, FBP) demonstrated robust transfer performance to all DLIR datasets, with AUC values in target datasets generally exceeding those in source datasets. In contrast, only models trained on the DLRL dataset exhibited non-inferior transferability when transferred to both AR50 and FBP datasets, with target AUCs being comparable to or higher than those in source datasets.
CONCLUSIONS: The transferability of radiomics models across CT reconstruction algorithms is asymmetric. Radiomics models demonstrated robust transferability from AR50 and FBP datasets to DLIR datasets for AP severity stratification.
PURPOSE: To investigate the transferability of radiomics models for acute pancreatitis (AP) severity stratification across CT images reconstructed with filtered back projection (FBP), adaptive statistical iterative reconstruction-Veo (ASIR-V), and deep learning image reconstruction (DLIR) algorithms.
METHODS: This retrospective study enrolled 79 AP patients who underwent contrast-enhanced CT. Five sets of images were reconstructed with FBP, ASIR-V at 50% blending level (AR50), and DLIR at low (DLRL), medium (DLRM), and high (DLRH) strength levels. A total of 837 radiomic features were extracted from the pancreatic parenchyma volume to generate five corresponding datasets. Radiomics models for stratifying AP severity were built on four machine learning algorithms, trained on three DLIR datasets (source datasets) and tested on AR50 and FBP datasets (target datasets). The source and target datasets were then swapped for reverse evaluation. Model transferability across datasets was assessed by comparing performance differences between source and target datasets, with area under the receiver operating characteristic curve (AUC) as the primary metric.
RESULTS: We identified a fundamental asymmetry in model transferability. Models trained on conventional reconstructions (AR50, FBP) demonstrated robust transfer performance to all DLIR datasets, with AUC values in target datasets generally exceeding those in source datasets. In contrast, only models trained on the DLRL dataset exhibited non-inferior transferability when transferred to both AR50 and FBP datasets, with target AUCs being comparable to or higher than those in source datasets.
CONCLUSIONS: The transferability of radiomics models across CT reconstruction algorithms is asymmetric. Radiomics models demonstrated robust transferability from AR50 and FBP datasets to DLIR datasets for AP severity stratification.