A dual stream architecture integrating pathology foundation models for automated tumor region segmentation in Ki-67 assessment

The quantification of the Ki-67 proliferation index is a cornerstone in breast cancer prognosis and treatment planning. While deep learning has shown promise, traditional convolutional neural networks (CNNs) lack global context, and standard vision transformers struggle with local boundary delineation. Recently, pathology-specific foundation models (FMs) like Phikon and UNI have emerged, offering robust feature representations. However, applying these patch-based models directly to pixel-wise segmentation remains challenging due to the loss of spatial resolution. In this work, we propose the dual-stream U-Net (DSU), a novel hybrid architecture designed to bridge the gap between the semantic power of foundation models and the spatial precision required for tumoral segmentation. Our approach integrates a frozen FM backbone (validating both Phikon and UNI ) with a parallel lightweight CNN encoder. This dual-stream design allows the model to leverage robust high-level semantic features from the FM while recovering fine-grained spatial details through the CNN stream. Furthermore, we employ a full-scale aggregation decoder to fuse multi-scale information dynamically. Experimental results using a leave-one-subject-out (cross-validation demonstrate that the DSU architecture achieves robust performance, outperforming traditional baselines (U-Net, TransUNet). Furthermore, incorporating boundary-sensitive metrics such as the 95th percentile Hausdorff distance confirms that DSU provides improved structural fidelity at the edges. While single-stream decoders can achieve comparable or higher volumetric overlap in some regimes, evaluations on independent test data reveal that Phikon-DSU and Uni-DSU provide improved spatial boundary delineation. These findings suggest that the proposed DSU framework offers a clinically viable preliminary proof-of-concept for leveraging foundation models in histology segmentation, serving as a first step toward automated Ki-67 scoring.