Analysis of digital pathology images with deep learning

Prostate1.png

Digital pathology is a new, rapidly expanding field in medical imaging. In digital pathology whole-slide scanners are used to digitize glass slides containing tissue specimens at high resolution (up to 160 nm per pixel), producing giga-pixel whole-slide images.

One of the principal tasks of pathologists is to detect cancer in histopathology images. Staining techniques are used to visualize parts of the tissue that are relevant for the diagnosis. As an example, hematoxilyn and eosin (H&E) staining is used to color cell nuclei in blue and connective tissue in pink. Immunohistochemistry (IHC) is a more advanced staining technique, used to highlight specific cells which exhibit, for instance, a function in the tumor growing mechanism. In all cases, the goal is to detect and characterize tumors.

Digitized histopathology images contain a huge amount of information, which represents a perfect source of data to develop deep learning algorithms for automatic image analysis. This is clearly demonstrated by the fact that in the Camelyon16 challenge, which we recently organized, the ten best scoring algorithms were all based on deep learning. In this challenge, we assessed the state of the art technology for solving the task of metastasis detection in sentinel lymph nodes.

Large databases of digital pathology images are available in our group, consisting of images of breast, colorectal, head and neck cancer, among others. Software libraries for handling gigapixel whole-slide images are available, as well as a workstation developed in house to visualize and annotate whole-slide images. Several projects are available for master students to work on deep learning for automatic analysis of digital pathology images (see Tasks). Deep learning algorithms will be developed based on the libraries Theano/Lasagne and TensorFlow.

Institute

The work will be executed in a high level environment for medical image analysis, at the Diagnostic Image Analysis Group (DIAG) of the Radboud University Medical Center. DIAG is a leading research group in computer-aided detection and diagnosis (CAD).

Tasks

Here are possible tasks suitable for your research project:

  • Develop a deep learning method to classify lung cancer sub-types in a large (>1,000) set of data.
  • Incremental deep learning: expand the knowledge of convolutional networks by adding new tissue types on-the-fly avoiding catastrophic forgetting.
  • Prediction of survival in lung cancer patients with deep learning
  • Prediction of DCIS progression to invasive cancer using deep learning based assessment of the stroma
  • Develop a deep learning method for automatic detection of metastases in IHC slides of sentinel lymph nodes
  • Develop a deep learning method for analysis of highly multiplexed immunofluorescence images (containing up to 7 independent channels)
  • Investigate the possibility of transferring the knowledge of a deep learning model developed for, e.g., breast cancer detection to colorectal cancer detection
  • Develop convolutional neural networks for fully automated segmentation of anatomic structures in biopsies from transplanted kidneys
  • Investigate the difference between using static or dynamic approaches to build training datasets of digital pathology images for convolutional networks
  • Build convolutional networks for the detection of nerves and vessels in colorectal cancer
  • Automated detection and classification of cell nuclei in colon histopathological images
  • Normalization of immune-histochemical signals in histopathology images

Requirements

  • Students with a major in computer science, biomedical engineering, artificial intelligence, physics, or a related area in the final stage of master level studies are invited to apply
  • Affinity with programming is required, interest and experience with machine learning and deep learning preferred

Information