Data-efficient pathology whole-slide image segmentation with slide-level conditioning

Whole-slide images (WSIs) are inherently high-resolution, gigapixel-scale data that require models to reconcile global slide context with fine-grained patch precision, yet progress is often limited by scarce pixel-level annotations. In this work, we address binary foreground segmentation of two histopathological target structures: glomeruli in PAS-stained kidney WSIs and atherosclerotic plaque lesions in murine aortic root sections, where the foreground corresponds to the dataset-specific structure of interest and all remaining tissue and background are treated as non-target regions. We introduce SlideFiLM, a segmentation framework that uses Prov-GigaPath whole-slide representations to guide tile decoding with slide context through a bounded slide-conditioned modulation operator that is shared across tiles. By injecting slide-level context across all tiles, the method achieves end-to-end segmentation on gigapixel WSIs. Across both benchmarks, SlideFiLM outperforms established baselines, with the largest gains for tissue structures that require global tissue context. Crucially, SlideFiLM maintains accuracy in low-data regimes, achieving robust performance under reduced training sets, and shows enhanced slide-level consistency as well as robustness across tissue types. These findings highlight SlideFiLM as an effective solution for WSI segmentation that incorporates global context, providing complete slide coverage, stable performance with limited labels, and consistent gains on structures that depend on broader contextual information.