EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

HierAnom‐Path: Zero‐Shot Rare Cancer Detection in Histopathology via Hierarchical Foundation Model Adaptation

Noora Saleem Jumaah, Mohanad Sameer Jabbar, Saif H. Alwaeli, Noor Ayesha, Mohamed Babiker A. Mohamed, Muhammad I. Khan

Computational and Systems Oncology · 2026-12 · 6 (1)

DOI: 10.1002/cso2.70022 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computational and Systems Oncology, online erschienen am 28.07.2026. Verfasst von Noora Saleem Jumaah und 5 weiteren Autor:innen. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Rare cancers collectively represent 20%–25% of all malignancies yet yield fewer than 50 annotated whole‐slide images (WSIs) per subtype in publicly available archives, rendering supervised deep learning inapplicable. We present HierAnom‐Path, to the best of our knowledge, the first framework for zero‐shot rare cancer anomaly detection in gigapixel histopathological WSIs that operates without any labeled rare cancer examples at any stage of training. HierAnom‐Path integrates four novel contributions: (1) a hierarchical two‐level BiomedCLIP adapter with cross‐cancer contrastive LoRA (Rank 16, 4.7 million trainable parameters) that learns morphological anomaly representations invariant across cancer subtypes without labeled rare cancer data; (2) RarePath‐Diffusion, a text‐conditioned latent diffusion model synthesizing photorealistic rare cancer pathology tiles from structured clinical prompts, providing training signal without any real labeled rare cancer slides; (3) the cross‐cancer‐subtype zero‐shot (CCSZ) benchmark protocol, the first evaluation framework enforcing strictly disjoint training and test cancer subtypes; and (4) a clinician‐aligned spatial anomaly heatmap with text‐grounded morphological explanations and a structured, auditable inference log. Evaluated on 30 rare brain tumor subtypes from the EBRAINS Digital Brain Tumour Atlas, HierAnom‐Path achieves a mean slide‐level AUROC of 0.841 ± 0.024, a +17.0 percentage point improvement over the strongest CLIP‐based baseline (AnomalyCLIP: 0.671 ± 0.031; DeLong test, p < 0.001, Bonferroni‐corrected), with tile‐level AUROC 0.783 ± 0.031 and AUPRO 0.762 ± 0.028. Multicenter validation on CAMELYON17 confirms scanner‐robust generalization (Slide‐AUROC 0.812 ± 0.019 across five independent sites; inter‐site range: 3.8 percentage points). Grade correlation on the PANDA prostate dataset ( n = 10,616 WSIs) yields Spearman ρ = 0.741 ( p < 0.001), and cross‐dataset evaluation on five non‐brain rare TCGA subtypes achieves a mean Slide‐AUROC of 0.798 ± 0.029. Blinded pathologist validation yields Fleiss' κ = 0.67 (substantial agreement) with explanation‐localization Dice 0.64 ± 0.11, confirming clinical interpretability. HierAnom‐Path closes the majority of the gap to a fully supervised upper bound (0.910 ± 0.018) while requiring zero rare cancer labels, establishing a new state of the art for zero‐shot rare cancer anomaly detection.

Zitieren

APA

Jumaah, N. S., Jabbar, M. S., Alwaeli, S. H., Ayesha, N., Mohamed, M. B. A., & Khan, M. I. (2026). HierAnom‐Path: Zero‐Shot Rare Cancer Detection in Histopathology via Hierarchical Foundation Model Adaptation. Computational and Systems Oncology, 6(1). https://doi.org/10.1002/cso2.70022

Vancouver

Jumaah NS, Jabbar MS, Alwaeli SH, Ayesha N, Mohamed MBA, Khan MI. HierAnom‐Path: Zero‐Shot Rare Cancer Detection in Histopathology via Hierarchical Foundation Model Adaptation. Computational and Systems Oncology. 2026;6(1). doi:10.1002/cso2.70022

BibTeX
@article{Jumaah2026,
  author = {Jumaah, Noora Saleem and Jabbar, Mohanad Sameer and Alwaeli, Saif H. and Ayesha, Noor and Mohamed, Mohamed Babiker A. and Khan, Muhammad I.},
  title = {HierAnom‐Path: Zero‐Shot Rare Cancer Detection in Histopathology via Hierarchical Foundation Model Adaptation},
  journal = {Computational and Systems Oncology},
  year = {2026},
  volume = {6},
  number = {1},
  doi = {10.1002/cso2.70022}
}
Kostenlos starten →