A Cross-Domain Deep Learning Prediction System for Ki-67 Proliferative Status in Breast Cancer Via Domain-Invariant Feature Fusion

Abstract

Ki-67 proliferative status is a clinically significant prognostic biomarker in breast cancer. It is routinely used to guide decisions about adjuvant chemotherapy and to provide prognostic information independent of tumour grade and nodal involvement. The current assessment relies on immunohistochemical analysis of biopsy or surgical specimens, a process that is invasive, susceptible to inter-observer variability, and poorly suited to longitudinal monitoring. Non-invasive estimation from full-field digital mammography offers a compelling alternative, since mammography already generates large volumes of routinely acquired data at no additional imaging cost. However, this approach is significantly hindered by systematic domain shift between imaging systems from different manufacturers, whose proprietary post-processing pipelines introduce vendor-specific differences in pixel intensity, contrast, and textural appearance. The present study proposes a patient-level prediction framework integrating shape-constrained radiomic features, attention-pooled multi-view deep representations, and domain-generalisation training objectives in order to address this issue. Shape-based radiomic descriptors are employed for their geometric invariance to vendor-specific processing, while an EfficientNet-B0 backbone augmented with MixStyle and Variance Risk Extrapolation is trained to produce acquisition-style-robust representations. A gated attention pooling mechanism aggregates features from four mammographic views into a single patient-level embedding, weighting projections adaptively according to their discriminative content. When evaluated under a leave-one-domain-out protocol across three scanner manufacturers on a cohort of $\mathbf{2, 4 1 0}$ patients, the fusion of deep and radiomic features outperformed either modality in isolation, achieving an AUC of 0.976 on the Hologic held-out set and 0.776 on Siemens, while GE remained the most challenging domain with a maximum AUC of 0.576. The results demonstrate that feature design is the primary determinant of cross-vendor generalisation and that geometrically invariant radiomics combined with domain-generalisation objectives provide a viable path towards vendor-agnostic mammographic Ki-67 biomarkers.