Mammographic radiomics and breast density for predicting PD-L1 expression in breast cancer
Yi-Shan Zhao, Hao Li, Can-Can Zhao, Yu-Heng Wang, Ping Wang, Zong-Yu Xie, Yu Ji, Hong LuCancer Imaging. 2026 Feb 7;26(1):38. doi: 10.1186/s40644-026-01001-3.
Abstract
Background: Programmed death-ligand 1 (PD-L1) expression is a critical biomarker for guiding immunotherapy in breast cancer, particularly in triple-negative subtypes. However, conventional assessments rely on invasive biopsies and are limited by tumor heterogeneity. This study aims to develop a non-invasive approach for predicting PD-L1 expression using mammography-based radiomics features integrated with clinicopathological variables and breast density, and to evaluate its performance in both an internal development cohort and an independent external validation cohort.
Methods: A total of 121 patients with breast cancer who underwent PD-L1 testing were retrospectively included, comprising 81 patients from Tianjin Medical University Cancer Institute & Hospital (development cohort, April 2023�September 2024) and 40 patients from the First Affiliated Hospital of Bengbu Medical University (external test cohort, January 2019�March 2025). Lesion regions of interest (ROIs) were manually annotated on both mediolateral oblique (MLO) and craniocaudal (CC) views for radiomic feature extraction using the SIMPACS Research platform. Additionally, standardized 1.5 cm � 1.5 cm ROIs were placed in the retroareolar parenchyma of both ipsilateral and contralateral breasts to evaluate background breast density. A multilayer perceptron (MLP) classifier was trained in the development cohort by combining lesion radiomic features, ipsilateral breast density radiomics, and clinicopathological variables, and then applied without recalibration to the external cohort. Performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.
Results: In the development cohort, the radiomic model incorporating clinicopathological information achieved an AUC of 0.610. When ipsilateral breast density was added, the AUC improved to 0.731. In contrast, the models with contralateral and bilateral density achieved lower AUCs of 0.535 and 0.537, respectively. In the independent external cohort, the final ipsilateral radiomics�clinical MLP model achieved an AUC of 0.629.
Conclusion: Mammography-based radiomics models may offer a non-invasive approach to predicting PD-L1 expression in breast cancer. The inclusion of ipsilateral breast density improves predictive performance and could support individualized immunotherapy decision-making. The observed performance in an external cohort provides preliminary evidence of cross-institutional generalizability, while highlighting the need for further optimization and validation in larger multicenter studies.