FTU-Seek: Foundation Model-Guided Hard-Negative Learning for Sparse Functional Tissue Unit Segmentation

Background/Objectives: Functional tissue units (FTUs), including tertiary lymphoid structures (TLSs), blood vessels, and glands, encode localized immune, vascular, and epithelial organization in histopathology. Accurate quantification of these structures is important for studying tissue architecture and disease-associated tissue organization. However, FTUs are frequently sparse, heterogeneous, and surrounded by large amounts of morphologically similar background tissue, making automated segmentation in whole-slide images (WSIs) challenging. We therefore developed FTU-Seek, a pathology foundation model-guided framework that treats morphology-aware negative-patch selection as a key component of sparse FTU segmentation. Methods: FTU-Seek uses frozen multi-depth features from the UNI pathology foundation model to train a patch-level classifier that distinguishes FTU-containing from FTU-absent tissue. Target-absent patches are subsequently ranked according to their predicted target-containing probabilities, and the highest-scoring hard negatives are selected through a static TopK strategy to construct compact segmentation training sets. The framework was evaluated using five-fold cross-validation and internal test cohorts across TLS, blood-vessel, and gland segmentation tasks, with an additional independent 30-WSI held-out cohort for TLS. Positive-only, all-tissue, random-negative, and matched random TopK sampling strategies served as comparators. Segmentation-derived phenotypes were further explored in external TCGA cohorts. Results: The patch-level classifiers achieved mean validation AUCs of 95.92%, 90.11%, and 98.16% for TLS, blood vessel, and gland classification, respectively. For TLS segmentation, the pre-specified Top1000 configuration retained 27.6% of the all-tissue training workload and achieved a slide-level Dice of 76.69 ± 11.89% on the independent 30-WSI held-out cohort. Compared with matched random Top1000 sampling, it improved Dice by 3.72 percentage points (95% CI, 2.05-5.38). Blood vessel and gland segmentation achieved performance approaching all-tissue training while reducing the retained training workload by approximately one-half and one-third, respectively. Compared with matched random sampling, classifier-guided hard-negative selection produced the greatest improvements for sparse and morphologically ambiguous FTUs. Exploratory TCGA analyses further showed associations of TLS phenotypes with overall survival, vascular phenotypes with overall survival and microvascular invasion, and glandular phenotypes with clinicopathological characteristics. Conclusions: FTU-Seek demonstrates that pathology foundation models can support sparse FTU segmentation not only through feature representation but also through morphology-aware construction of segmentation training sets. By prioritizing informative hard negatives, the framework reduces redundant segmentation-training workload while maintaining competitive segmentation performance and supporting quantitative tissue phenotyping from routine histopathology.