EUVIMEDEUVIMED entdecken
Crossref

Interpretable Tissue-Composition Biomarkers from Deep Semantic Segmentation for Molecular Subtype and Survival Prediction in Breast Cancer

B Mahesh, Geethanjali Nellore

Research Digest on Engineering Management and Social Innovations · 2026

DOI: 10.46647/rdems0205077 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Research Digest on Engineering Management and Social Innovations, 2026. Verfasst von B Mahesh und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

Whole-slide image (WSI) classifiers can predict outcomes in breast cancer, but their decisions are difficult to interpret, whereas expert-annotated tissue segmentation datasets are rarely connected to patient outcomes.

Objective

We investigate whether tissue-composition and spatial-architecture features derived from deep semantic segmentation, trained on the Breast Cancer Semantic Segmentation (BCSS) dataset, can predict molecular subtype and survival in the TCGA-BRCA cohort, and how they compare with end-to-end multiple-instance learning (MIL).

Methods

We propose a three-stage framework: (i) stain-normalised tissue segmentation using convolutional, transformer-based and foundation-model-encoder architectures with patient-level cross-validation; (ii) extraction of 42 interpretable compositional and spatial features; and (iii) subtype classification and survival modelling with regularised Cox, random survival forests and gradient boosting, benchmarked against ABMIL and TransMIL on pathology foundation-model embeddings.

Results

The best segmenter (UNI-UperNet) reached a Dice score of 0.81 and mIoU of 0.69. Segmentation-derived features discriminated PAM50 subtypes with macro-AUC 0.84 (Basal-like AUC 0.92, Luminal B AUC 0.78). For overall survival, the feature-based model combined with age and stage achieved a C-index of 0.75 compared with 0.66 for clinical variables alone and 0.72 for the best end-to-end model (TransMIL (UNI)); high- versus low-risk groups differed in survival (HR 2.60, 95% CI 1.70–4.00, log-rank p < 0.001). The most influential features were TIL fraction in stroma, tumour-stroma ratio and stage.

Conclusion

Segmentation-derived biomarkers provide transparent prognostic information that is competitive with end-to-end models, supporting their use as auditable components of computational-pathology pipelines.

Zitieren

APA

Mahesh, B., & Nellore, G. (2026). Interpretable Tissue-Composition Biomarkers from Deep Semantic Segmentation for Molecular Subtype and Survival Prediction in Breast Cancer. Research Digest on Engineering Management and Social Innovations. https://doi.org/10.46647/rdems0205077

Vancouver

Mahesh B, Nellore G. Interpretable Tissue-Composition Biomarkers from Deep Semantic Segmentation for Molecular Subtype and Survival Prediction in Breast Cancer. Research Digest on Engineering Management and Social Innovations. 2026. doi:10.46647/rdems0205077

BibTeX
@article{Mahesh2026,
  author = {Mahesh, B and Nellore, Geethanjali},
  title = {Interpretable Tissue-Composition Biomarkers from Deep Semantic Segmentation for Molecular Subtype and Survival Prediction in Breast Cancer},
  journal = {Research Digest on Engineering Management and Social Innovations},
  year = {2026},
  doi = {10.46647/rdems0205077}
}
Kostenlos starten →