Deep learning and digital pathology

Digital pathology has recently become widely accepted due to advances in technology and regulations. Essentially, in digital pathology, a scanner creates a digital copy of the traditional glass slide that can be stored on a local or cloud-based server and processed anywhere with a computer and an internet connection. Its use for primary diagnosis has revolutionised pathology and is shaping the future of the discipline.

The DigiPath team deals with various challenges in the field of digital pathology. These include, for example, the classification of tumor and non-tumor tissue or of different tumor types based on digital microscopy images. This task would normally require extensive microscopic evaluation by pathologists. In our team, we develop powerful and robust methods for this purpose using both deep learning and classical machine learning techniques.

Research areas

  • Semantic segmentation
  • Interpretability/explainability of Unet-like neural networks
  • Stain normalization with GANs
  • Normalizing flows and invertible UNet for synthetic data generation
  • Training with weak labels + differential diagnosis

Leader

Bild Daniel Otero Baguer

Dr. Daniel Otero Baguer

Deep Learning and Digital Pathology

DigiPath-Viewer

The digipath team is also developing a "DigiPath Viewer", a web application that can be used to annotate digital microscopy images and display the results of the models. The "DigiPath Viewer" will visualize our results and also simplify the work of pathologists. Some of the main features are:

  • Integration with stylus pen for natural and precise annotations.
  • A semi-automatic annotation tool that speeds up the annotation process. The annotation polygon can be easily adjusted by dragging/deleting points.

  • Visualization of automatic artificial intelligence tumor predictions as overlays on the original images. The transparency of the overlay as well as the threshold to distinguish between tumor or normal tissue can be easily adjusted.

CAMELYON 17 Challenge

Participation in Challenges

Our team placed 9th in the unofficial leaderboard of the CAMELYON 17 Challenge, which remained open after the official submission deadlines and received more than 100 submissions from the best companies and research groups in the field of digital pathology. The goal of this challenge is to develop a method for automated detection and classification of breast cancer metastases in whole-slide images of histological lymph node sections.

 

Team

Bild Rudolf Herdt

Rudolf Herdt

Deep Learning and Inverse Problems

Bild Jean Le'Clerc Arrastia

Jean Le'Clerc Arrastia

Deep Learning and Digital Pathology

Projects

Visualisierung von KI-Ergebnissen zur Krebserkennung

SPAplus - Small Data problems in digital pathology

BMBF-Projekt
Duration: 01.04.2020 - 31.03.2023
PI: Peter Maaß

In SPAplus, mathematically based methods for data augmentation via deep learning are being developed for digital pathology. In addition, concepts for the training of mathematical data analysts as well as information and networking events will be developed and implemented in program-accompanying measures.

DIAMANT Abbildung

DIAMANT - Digital image analysis and imaging mass spectrometry for differentiation of non-small cell lung cancer

BMBF-Projekt
Duration: 01.01.2020 - 31.12.2022
PI: Peter Maaß

In DIAMANT, molecular information from MALDI IMS data is combined with detailed anatomical information from digital microscopy images. Using an integrated analysis of data from both complementary modalities, a classification model for cancer diagnosis is being developed that is significantly more accurate than existing models based on only one of the two modalities.

Logo MALDISTAR

MALDISTAR - Study on quality assessment, standardization and reproducibility of MALDI imaging mass spectrometry data

KTS-Projekt
Duration: 01.07.2019 - 30.06.2022
PI: Peter Maaß

MALDI imaging represents an established method for the spatial study of biomolecules. However, despite many advantages, it is becoming increasingly clear that the data are subject to high variability. For this reason, quality assessment tools and new calibration and cross-normalization methods are being developed in MALDISTAR.

Publications

J. Le Clerc Arrastia, N. Heilenkötter, D. Otero Baguer, L. Hauberg-Lotte, T. Boskamp, S. Hetzer, N. Duschner , J. Schaller , P. Maaß.
Deeply Supervised UNet for Semantic Segmentation to Assist Dermatopathological Assessment of Basal Cell Carcinoma.
MDPI Journal of Imaging, 71 7(4), Meisenbach Verlag, Bamberg, 2021.

DOI: https://doi.org/10.3390/jimaging7040071


J. Behrmann, C. Etmann, T. Boskamp, R. Casadonte, J. Kriegsmann, P. Maaß.
Deep Learning for Tumor Classification in Imaging Mass Spectrometry.
Bioinformatics, 34(7):1215-1223, Oxford University Press, 2018.

DOI: 10.1093/bioinformatics/btx724

 

 

C. Etmann, M. Schmidt, J. Behrmann, T. Boskamp, L. Hauberg-Lotte, A. Peter, R. Casadonte, J. Kriegsmann, P. Maaß.
Deep Relevance Regularization: Interpretable and Robust Tumor Typing of Imaging Mass Spectrometry Data.
Zur Veröffentlichung eingereicht.