Annals of surgical oncology · 2026
AI-FLEET: Phase I-Multimodal Deep Learning Model for Phyllodes Tumor Classification.
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.
Holt Logan, Chamberlain Victoria, Shern Tyler, Abir Farhan Fuad, Daly Abigail, Anderman Kyle, Gadd Michele A, Verdial Francys C, Kwait Rebecca M, Specht Michelle C, Smith Barbara L, Brattain Laura, Ozmen Tolga
Аннотация
BACKGROUND: Fibroepithelial breast lesions, including fibroadenomas and phyllodes tumors (PTs), can be difficult to classify on needle biopsy. Misclassification may result in unnecessary excisions of benign fibroadenomas or delays and repeat operations for borderline/malignant PTs. Artificial Intelligence for Fibroepithelial Lesion Evaluation and Extrication Technology (AI-FLEET) is a multi-stage program designed to improve diagnostic accuracy and reduce inconclusive preoperative assessments by integrating radiologic, pathologic, and clinical data.
PATIENTS AND METHODS: In this first phase, we retrospectively analyzed patients with histologically confirmed PTs. Borderline and malignant PTs were grouped together owing to similarities in margin management and the limited number of cases. Models were trained to distinguish benign from borderline/malignant PTs using ultrasound images and clinical variables (age, body mass index (BMI), race/ethnicity, menopausal status, echogenicity, and tumor size). Multiple convolutional and attention-based encoders were evaluated using subject-stratified five-fold cross-validation.
RESULTS: The cohort included 81 patients (65 benign, 16 borderline/malignant PTs) with 1638 ultrasound images. The multimodal ConvNeXt model achieved an accuracy of 0.91 (AUC 0.94), while the multimodal ResNet18 achieved an accuracy of 0.92 (AUC 0.94). Other multimodal architectures showed lower performance. Ultrasound-only and clinical-only models reached AUCs of 0.89 and 0.78, respectively. Saliency analyses identified intratumoral heterogeneity as an important predictive feature.
CONCLUSIONS: Multimodal deep learning models combining ultrasound and clinical factors achieved high accuracy in differentiating benign from borderline/malignant PTs, demonstrating the feasibility of AI-assisted assessment of fibroepithelial lesions. Phase II will expand this work by incorporating histopathology and fibroadenoma cases to further enhance radiologic-pathologic integration.
PATIENTS AND METHODS: In this first phase, we retrospectively analyzed patients with histologically confirmed PTs. Borderline and malignant PTs were grouped together owing to similarities in margin management and the limited number of cases. Models were trained to distinguish benign from borderline/malignant PTs using ultrasound images and clinical variables (age, body mass index (BMI), race/ethnicity, menopausal status, echogenicity, and tumor size). Multiple convolutional and attention-based encoders were evaluated using subject-stratified five-fold cross-validation.
RESULTS: The cohort included 81 patients (65 benign, 16 borderline/malignant PTs) with 1638 ultrasound images. The multimodal ConvNeXt model achieved an accuracy of 0.91 (AUC 0.94), while the multimodal ResNet18 achieved an accuracy of 0.92 (AUC 0.94). Other multimodal architectures showed lower performance. Ultrasound-only and clinical-only models reached AUCs of 0.89 and 0.78, respectively. Saliency analyses identified intratumoral heterogeneity as an important predictive feature.
CONCLUSIONS: Multimodal deep learning models combining ultrasound and clinical factors achieved high accuracy in differentiating benign from borderline/malignant PTs, demonstrating the feasibility of AI-assisted assessment of fibroepithelial lesions. Phase II will expand this work by incorporating histopathology and fibroadenoma cases to further enhance radiologic-pathologic integration.