Digital Pathology

Tumor Budding

© Fraunhofer IIS
© Fraunhofer IIS

Tumor budding is an independent predictor of lymph node metastases in pT1 colorectal cancer and predictor of survival in stage II colorectal cancer1.   

Goal: automatic scoring

Partners: joint effort with the Institute of Pathology of the University Hospital Erlangen

Current Status:

  • developed automatic analytics App for an immunohistochemistry staining (pan-cytokeratin – PCK) that is able
  • to detect tumor invasion front
  • to detect tumor buds
  • to detect hot spots and determine low/intermediate/high budding based on cut offs
  • algorithm trained on 40x PCK-stained whole-slides

Database: over 100 whole-slides with over 50.000 annotated tumor buds

Solution:

  1. Detect tumor bud candidates using classical image processing that evaluates size, color, and distance to tumor
  2. Weed out false positives using a Convolutional Neural Network trained to classify tumor buds

------------------------------------------------------------------

1Lugli et al, Recommendations for reporting tumor budding in colorectal cancer based on the International Tumor Budding Consensus Conference (ITBCC) 2016

 

Goal: partition different tissue types and detect adeno carcinoma in H&E stained colon tissue sections

Partners: joint effort with the Institute of Pathology of the University Hospital Erlangen

Current Status:

  • Created annotated database of 20x H&E slides
  • Currently training AI models to classify patches from Whole-Slide
  • Currently rescanning slides with other scanners to become more robust to scanner variations

Database:

  • over 300 whole-slides that amount to nearly 30 mio. tiles of size 227x227
  • many classes: tumor, tumor cell clusters, fat, mucosa, connective tissue, blood vessels, lymphoid follicles

Solution:

  • end to end training of Convolutional Neural Network

PathoScan Project: Automated Digitalization in Routine Pathology

PathoScan: Artificial Intelligence in routine pathology
© ©Vshyukova - stock.adobe.com
PathoScan – Optimization of morphological work processes through an automated digitized pathology system

Goal: Digital and automated microscopy workflow including pre-scan QC using an overview camera

Partner: PreciPoint GmbH, University of Regensburg, Technical University of Munich TUM, HTI bio-X GmbH

             

Current Status:

  • Systematic collection and prioritization of possible error artifacts that can occur during sample preparation
  • Collection and digitization of problematic slides

Database: Slides are currently being collected in Regensburg and Munich and then digitized. Using the database, the scientists at Fraunhofer IIS will develop and validate the image analysis algorithms.

Our Solution:

  • Coloring (HTI Smart Automation) and digitization (PreciPoint) using a fully automated digitization workflow
  • Automatic image analysis using the overview camera to detect various quality problems such as protruding cover glasses, air bubbles, foreign particles, folds, cracks
  • Detection of over / under-staining or faulty staining by objective quantification of the coloring
  • Cost and time efficient quality assurance procedures

Low-Budget Diagnostics of Gastroesophageal Tumors

DigImmune
© Fraunhofer IIS

Goal: histological image analysis – comparison iSTIX vs whole-slide-scanner

Partners: joint effort with the Definiens AG, Institute of Pathology of the University Hospital Erlangen, Institute of Pathology of Bayreuth Hospital

Current Status:

  • Creation of a comprehensive annotated database: data from various clinical centers, different scanners and staining methods (H&E plus 9x immunohistochemistry), both of resections and of biopsies
  • Quantification of IHC samples and analysis of correlations between IHC biomarkers and between patients

Database: completed

  • A total of 1000 slides (100 cases with 10 stains each) from Bayreuth and Erlangen digitized with 3DHISTECH scanners
  • Additional overview scan (50x) of the H&E staining of all cases manually scanned with iSTIX
  • Additionally, tumor region selected by the pathologist in H&E staining with iSTIX in high resolution (200x) manually scanned

Solution:

  • Scan whole-slides with Scanners
  • Scan tumor regions with iSTIX
  • Image analysis of whole-slides created with iSTIX vs. scanner

 

"DigImmune (Digitale Diagnostik für die Immuntherapie von Krebspatienten)"

Single Cell Tumor Detection in Lymph Nodes

© Fraunhofer IIS

Goal: early detection of cancer metastasis by detecting even single disseminated tumor cells in a lymph node.

Partners: joint effort with the Fraunhofer ITEM and Fraunhofer IPA

Current Status:

  • Development of single components (tissue grinder, scanning and image analysis, single cell DNA amplification, laboratory notebook software) finished
  • Evaluation ongoing

Database:

  • ~50 slides of melanoma and lung cancer

Solution:

  • Prior work by University Tübingen and Fraunhofer ITEM has shown that even a presence of only a handful of disseminated cancer cells per million lymphocytes has a negative impact on survival
  • Fraunhofer IPA has developed TissueGrinder for dissolving a lymph node into a cell suspension
  • Immunocytological staining using melanoma marker (gp100)
  • Scanning of cell suspension with Fraunhofer IIS’s SCube scanning platform
  • Fraunhofer IIS image analysis: detection of tumor cell candidates and tumor vs. artefact classification with Convolutional Neural Network
  • Molecular single cell analysis with assay developed by Fraunhofer ITEM (commercially available as Ampli1TM WGA)

Seeking Commercialization Partners


-----------------------------------------------------------------

Further Reading:
1. Ulmer A., et al., Quantitative measurement of melanoma spread in sentinel lymph nodes and survival. PLoS Med, 2014. 11(2): e1001604.
2. Ulmer A., et al., The sentinel lymph node spread determines quantitatively melanoma seeding to non-sentinel lymph nodes and survival. Eur J Cancer, 2018. 91: p. 1-10
3. Werner-Klein M., et al. Genetic alterations driving metastatic colony formation are acquired outside of the primary tumour in melanoma. Nat Commun, 2018. 9(1): p. 595

xAI – Explainable AI in Digital Pathology

© BMBF

Goal: turn AI algorithms used in Digital Pathology from a black box into a glass box

Partners: University of Bamberg (Cognitive Systems, Prof. Dr. Ute Schmid), Fraunhofer HHI

Current Status:

  • Research and evaluation of technologies to make AI algorithms more explainable

Solution:

  • Display results from neural networks in an easily comprehensible and transparent manner
  • Take into account uncertainties from the annotated training data as well from the model itself

Funded by Federal Ministry of Education and Research (BMBF)
Transparent Medical Expert Companion (TraMeExCo)

---------------------------------------------------------

Literature:
Schmid, Wittenberg (2019) TraMeExCo - Medical Expert Companion (Vortrag). All-Hands-Meeting (5. Juni 2019, Dortmund), Förderschwerpunkt Maschinelles Lernen des BMBF. -- [Slides], [Poster]
Rieger, Finzel, Seuß, Wittenberg, Schmid (2019) Make Pain Estimation Transparent: A Roadmap to Fuse Bayesian Deep Learning and Inductive Logic Programming (Poster). 41st Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC, 23.-27. Juli 2019, Berlin)

Bruns, Volker, Geppert, Carol (2019), Einsatz von Künstlicher Intelligenz in der digitalen Pathologie - Auf dem Sprung in die Routine?, Trillium Diagnostik 2/2019, https://www.trillium.de/zeitschriften/trillium-diagnostik/ausgaben-2019/td-heft-22019/pathologie/einsatz-von-kuenstlicher-intelligenz-in-der-digitalen-pathologie.html

Development of Robust AI systems for Digital Pathology

Goal: design AI algorithms that are robust to real-world heterogeneity (different scanners, staining variances, sub-optimal slide quality)

Current Status:

  • Developed algorithms for simulating staining variances
  • Creating multi-scanner database of H&E stained colon section
  • Developing Prototypical Networks (PN) for Few-Shot Learning: this will enable the development of new classifiers using only little data and few annotations

Database

  • In development

Solution:

  • Represent real-world heterogeneity in training database or normalize data to match training data characteristics

Research Project: ADA Center – Applications Data Analytics
Subproject: Explainable AI in Life Science and Automotive