Academic radiology · 2026
Ultrasound-Based Habitat Radiomics for Differentiating Phyllodes Tumor from Breast Fibroadenoma.
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.
Xue Qinwen, Ding Wenzhen, Lu Zhonglin, Chen Fang, Jia XiaoHong, Liu Jie, Sun Jiahong, Li Shuao, Zhou Jianqiao, Zhao Cheng, Yu Jie, Liang Ping
Аннотация
RATIONALE AND OBJECTIVES: Phyllodes tumors (PTs) and fibroadenomas (FAs) have overlapping ultrasound (US) appearances. This study aimed to develop and evaluate an interpretable machine-learning model integrating US radiomics, habitat features, and clinical variables.
MATERIALS AND METHODS: This retrospective multicenter study included 1950 pathologically confirmed lesions (1029 FAs and 921 PTs) divided into training (n=918), validation (n=474), and test (n=558) cohorts. Nineteen voxelwise features were used for K-means habitat segmentation. The number of clusters was selected from K=2-10 in the training cohort and evaluated across 50 random seeds. Whole-lesion, habitat, and combined habitat-radiomics (HR) signatures were developed using a train-only feature-selection pipeline, and ExtraTrees was selected as the classifier. The final combined model was built by probability-level late fusion of the ExtraTrees HR model and a logistic clinical model based on age, lesion size, and orientation.
RESULTS: K=3 yielded the highest Calinski-Harabasz index (200,927.01). Clustering was stable across 50 random seeds, with mean adjusted Rand indices of 0.9942, 0.9948, and 0.9941 in the training, validation, and test cohorts, respectively. Feature selection retained 39 whole-lesion features and 34 habitat features (habitats 1-3: 1, 14, and 19). The combined model achieved AUCs of 0.921 (95% CI, 0.904-0.939), 0.797 (95% CI, 0.754-0.835), and 0.844 (95% CI, 0.813-0.877) in the training out-of-fold, validation, and test cohorts.
CONCLUSION: In this multicenter cohort, a model integrating US habitat radiomics with age, lesion size, and orientation differentiated PTs from FAs with consistent discrimination across data partitions. Shapley additive explanations analysis provided global and case-level explanations of the combined prediction.
MATERIALS AND METHODS: This retrospective multicenter study included 1950 pathologically confirmed lesions (1029 FAs and 921 PTs) divided into training (n=918), validation (n=474), and test (n=558) cohorts. Nineteen voxelwise features were used for K-means habitat segmentation. The number of clusters was selected from K=2-10 in the training cohort and evaluated across 50 random seeds. Whole-lesion, habitat, and combined habitat-radiomics (HR) signatures were developed using a train-only feature-selection pipeline, and ExtraTrees was selected as the classifier. The final combined model was built by probability-level late fusion of the ExtraTrees HR model and a logistic clinical model based on age, lesion size, and orientation.
RESULTS: K=3 yielded the highest Calinski-Harabasz index (200,927.01). Clustering was stable across 50 random seeds, with mean adjusted Rand indices of 0.9942, 0.9948, and 0.9941 in the training, validation, and test cohorts, respectively. Feature selection retained 39 whole-lesion features and 34 habitat features (habitats 1-3: 1, 14, and 19). The combined model achieved AUCs of 0.921 (95% CI, 0.904-0.939), 0.797 (95% CI, 0.754-0.835), and 0.844 (95% CI, 0.813-0.877) in the training out-of-fold, validation, and test cohorts.
CONCLUSION: In this multicenter cohort, a model integrating US habitat radiomics with age, lesion size, and orientation differentiated PTs from FAs with consistent discrimination across data partitions. Shapley additive explanations analysis provided global and case-level explanations of the combined prediction.