Abstract PS3-06-30: Mammoscope: a clinically-informed foundation model for high-resolution mammography interpretation

Abstract

Abstract Artificial intelligence (AI) models in breast imaging have shown promise in automating classification tasks but often lack the nuanced decision-making radiologists apply in clinical practice. We developed a high-resolution vision transformer-based foundation model trained on over 650,000 mammograms from 14 public datasets and validated on internal cohorts. The model, termed MammoScope, incorporates radiologist-informed approaches such as bilateral comparison, priors, and lesion-context reasoning using image registration and multiview analysis. Self-supervised pretraining was performed using the DINO framework for improved performance, followed by fine-tuning on downstream tasks including benign versus malignant classification, BI-RADS prediction, and lesion subtype stratification. Citation Format: C. Sadée, C. Lin, D. Raymond, P. Shah, K. Sangani, Q. Xu, N. Hundal, A. Chun, T. Onyemeh, Z. Onah, A. Ilo, K. Hartmann, B. Dashevsky, I. Okoye, E. McDonald, O. Gevaert. Mammoscope: a clinically-informed foundation model for high-resolution mammography interpretation [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-06-30.

Ähnliche Arbeiten