Meaningful Digital Biomarkers

Challenge: High Information Density

• Histopathology images carry high information density.
• To answer clinically relevant questions, meaningful quantitative biomarkers need to be extracted.
• Predictive information can only be obtained from the complex interplay of different molecular mechanisms via robust algorithms.
   

Our Approach: Develop Meaningful Digital Biomarkers

• Using advanced artificial intelligence (AI) technology, we derive novel digital biomarkers from tissue sections that capture previously hidden clinically relevant information.
• Digital biomarkers characterizing tissue morphology can help make better use of routinely available tumor tissue to improve disease prognosis and understand disease mechanisms.
• Digital biomarkers that characterize the immune contexture in the tumor microenvironment can help identify subgroups of patients who will benefit most from personalized therapies.
• Through rigorous validation, we ensure that data-driven digital biomarkers are clinically meaningful.
• Fraunhofer MEVIS has more than 10 years of experience in developing robust image analysis solutions that support pathologists in the assessment of biomarkers, reducing workload and simplifying workflows.

Data-driven discovery of image-based biomarkers

• Classical, knowledge-driven biomarker research is complemented by data-driven discovery of biomarkers to identify prognostic features from the combination of images and tabular clinical data.
• We have implemented an identification of candidate biomarkers based on cell classes density features ranked by their potential for the predictive stratification [1].
• Deep neural networks can identify complex patterns without using hand-crafted features and manual annotations, with foundation models serving as powerful feature extractors [2], [3].
• Relevant corresponding image features can be discovered by analyzing the learned model using, e.g., explainability methods like attention weight visualization
   

References

[1] Data-Driven Discovery of Immune Contexture Biomarkers. Schwen et al. 2018. article

[2] Overcoming data scarcity in biomedical imaging with a foundational multi-task model. Schäfer et al. 2024. article

[3] Tissue concepts: Supervised foundation models in computational pathology. Nicke et al. 2025. article

Physiology-Based Quantification

• Many image algorithms use square tiles for quantification, which does not reflect physiological structures.
• Reflecting cellular properties depending on their location and interactions between neighboring cells, we
  • Compute biomarkers based on distances between cells [1],
  • Assess the tumor microenvironment by quantifying cell distances depending on the distance from the tumor boundary [1], and
  • Quantify tissue properties in biologically meaningful regions such as liver lobules [2, 3, 4].
     

References

[1] Data-Driven Discovery of Immune Contexture Biomarkers. Schwen et al. 2018. article

[2] Zonated quantification of steatosis in an entire mouse liver. Schwen et al. 2016. article

[3] Automated Detection of Portal Fields and Central Veins in Whole-Slide Images of Liver Tissue. Budelmann et al. 2022. article

[4] Joint zonated quantification of multiple parameters in hepatic lobules. Arp Laue et al. 2025. article

Quantifying Chromosomal Aberrations with FISH

• Specific chromosomal abnormalities like gene multiplications or translocations can be made visible with fluorescence in situ hybridization (FISH)
• We have developed solutions for automated detection and quantification of such abormalities by
  • Segmenting cell nuclei in 4′,6-diamidino-2-phenylindole (DAPI)
  • Detecting FISH signals and analyzing their patterns within nuclei bounds
  • Quantifying FISH signals even for large gene multiplication counts (e.g., ERBB2)

References

[1] Automated density-based counting of FISH amplification signals for HER2 status assessment. Höfener et al. 2019. article

Representative Datasets

• Digital biomarker studies are commonly based on non-representative datasets with narrowly defined patient collectives and strictly controlled environmental conditions
• We explore digital biomarkers based on cancer registry data covering the full spectrum of routine clinical cases to uncover meaningful disease patterns [1].
• We have extensive expertise in how to collect sufficient data and avoid bias so that datasets are truly representative [2].

References

 [1] Zusammenführung von Krebsregisterdaten und multimodalen, melderbasierten Diagnostikdaten zur KI-basierten Biomarker-Detektion. CanConnect

 [2] Recommendations on compiling test datasets for evaluating artificial intelligence solutions in pathology. Homeyer et al. 2022. article

Focused Scoring

• Spatial heterogeneity of tissue can make quantitative assessment unreliable and a robust discrimination between different groups difficult.
• Focused scoring is a hotspot analysis based on percentiles of scores obtained in a tile-based manner [1].
• Focused scoring offers robustness with respect to imaging and detection artifacts and helps to reduce sample sizes for showing significant results.

Reference

[1] Focused scores enable reliable discrimination of small differences in steatosis, Homeyer et al 2018. article

Customized Solutions for Estabished Biomarkers

• There are a large number of established histological biomarkers based on the quantification of cellular structures or the classification of morphological tissue patterns that benefit from computer support
• We can leverage extensive expertise and software components to quickly develop customized new solutions for biomarker assessment [1, 2, 3]
• We attach great importance to creating solutions that are both robust and efficient as well as easy to maintain and operate [4, 5]

References

[1] The NCI Imaging Data Commons as a platform for reproducible research in computational pathology. Schacherer et al. 2023. article

[2] Automated density-based counting of FISH amplification signals for HER2 status assessment. Höfener et al. 2019. article

[3] Deep learning nuclei detection: A simple approach can deliver state-of-the-art results.  Höfener et al. 2018. article

[4] Automated quantification of steatosis: agreement with stereological point counting. Homeyer et al. 2017. article

[5] Evaluation of the Transferability of a New “Learning” Histologic Image Analysis Application. Arlt et al. 2016. article