Related Experiment Video
Updated: Sep 2, 2026
Quantitation of Protein Expression and Co-localization Using Multiplexed Immuno-histochemical Staining and Multispectral Imaging
Published on: April 8, 2016
Sample-Efficient Transfer Learning for Histopathological Detection of HNSCC
Till Nicke1, Fleming Wolf-Eli Michelsen2, Jan-Raphael Schäfer1
1Fraunhofer Institute for Digital Medicine MEVIS, Bremen, Germany.
Background:
Head and neck squamous cell carcinoma (HNSCC) is a common malignancy with increasing incidence. Histopathological assessment remains the diagnostic gold standard, but is labor-intensive and affected by a growing shortage of specialized pathologists. Limited availability of annotated data poses a major challenge for AI-based support tools in digital pathology.
Methods:
We evaluated the transfer performance of a domain-specific foundation model, Tissue Concepts, compared to standard ImageNet pretraining for HNSCC detection in histopathological whole slide images. A multicentric, retrospective dataset comprising digitized H&E-stained slides from The Cancer Imaging Archive was used (390 slides, 268 HNSCC cases, 112 patients). Tumor regions were annotated by expert pathologists. Model representations were assessed without task-specific fine-tuning using a standardized classifier probe. Sample efficiency was investigated on patch-based and patient-based levels to estimate the number of annotated cases required for robust performance.
Results:
The foundation model consistently outperformed ImageNet pretraining across all experiments. On a patch-based level, Tissue Concepts achieved substantially higher discrimination performance with minimal training data (AUROC = 0.70 vs. = 0.50 using only 1 patch). On a patient-based level, the foundation model reached robust performance with a limited number of annotated cases (AUROC = 0.90 with 10 cases), whereas ImageNet-based representations showed slightly inferior performance despite substantially larger training sets (AUROC = 0.87 with 50 cases). Performance improvements plateaued beyond 10 cases, indicating diminishing returns with additional annotations.
Conclusion:
Domain-specific foundation models enable competitive HNSCC detection under annotation-scarce conditions. However, performance depends on the representativeness of annotated cases and tumor heterogeneity across anatomical sites and institutions. Further multicenter and prospective studies are required to assess generalizability and integration into routine histopathological workflows.
