Back to 2026 Abstracts
AI-FLEET: Phase II- Multimodal Deep Learning for Fibroepithelial Breast Lesion Classification
Logan R. Holt BS1, Victoria Chamberlain BS1, Tyler P. Shern BA1, Farhan Abir MS2, Michele Gadd MD1, Francys C. Verdial MD MPH1, Michelle Specht MD1, Barbara L. Smith MD PhD1, Laura Brattain PhD3, Tolga Ozmen MD1 1Division of Breast Surgery, Mass General Brigham, Boston, MA, USA. 2Department of Electrical and Computer Engineering, University of Central Florida, Orlando, FL, USA. 3Department of Internal Medicine, University of Central Florida College of Medicine, Orlando, FL, USA.
Background: Fibroepithelial lesions (FELs) are the most common breast tumors in young women, ranging from benign fibroadenomas (FAs) to phyllodes tumors (PTs). In 15% of cases, needle biopsy cannot distinguish FAs from PTs, necessitating surgical excision for many benign FAs. Artificial Intelligence for Fibroepithelial Lesion Evaluation and Extrication Technology (AI-FLEET) Phase II aimed to differentiate FAs and PTs preoperatively. Methods: We curated a single institution retrospective database of FELs excised between 2003-2024. Multimodal deep learning model inputs included ultrasound images and clinical features (age, echogenicity, and tumor size). We utilized the Ultrasound Foundation Model (USFM), and multiple instance learning enabled classification using all ultrasound images from each patient. A multimodal model was created using late fusion, in which the outputs of the unimodal models were combined using logistic regression. Model performance was assessed using subject-stratified 5-fold cross-validation (80% training, 20% validation). Results: 688 patients (409 FAs and 279 PTs) with 4,063 ultrasound images were analyzed. The multimodal model performed better than imaging-only and clinical-only models, achieving the highest AUC 0.775 [0.741, 0.811], accuracy 0.701 [0.671, 0.733], F1-score 0.634 [0.591, 0.680], PPV 0.672 [0.629, 0.721], NPV 0.760 [0.728, 0.794], sensitivity 0.638 [0.587, 0.692], and specificity 0.745 [0.709, 0.781]. Conclusion: The multimodal model outperformed imaging-only and clinical-only approaches in differentiating FAs from PTs. Future iterations will incorporate digitized pathology slides to enhance preoperative differentiation of FELs.
Back to 2026 Abstracts
