This is an AI for Health MSc project. Students are elgible to receive a monthly reimbursement of €500,- for a period of six months. For more information please read the requirements.
Clinical Problem
In medical imaging, data acquisition conditions differ among institutions, which leads to variations in image properties, known as domain shift. Convolutional neural networks (CNN) are sensitive to domain shifts, which can result in poor generalization. Stain variation in histopathological slides is a prominent example. A number of methods have been proposed to tackle domain shift: stain normalization, domain augmentation, domain adversarial training, classical data augmentation and domain adaptation. Each of the aforementioned methods improves CNNs performance on data from external centers to a certain extent. At the same time, none of these methods guarantees stain invariance.
Solution
In this project we want to make a step towards stain invariant models (trained only on data from a single institution which generalize well to data from other (new) institutions). Group Convolutional Neural Networks (GCNN) have been applied to introduce rotation and translation invariance in histopathology. In M. Lafarge et. al, (2021) models learn feature representations with a discretized orientation dimension that guarantees that their outputs are invariant under a discrete set of rotations. In this project we plan to use GCNNs to encourage and guarantee stain invariance, at least under certain constraints. We first plan to define a symmetry group in one of the color spaces. Secondly, we will define functions that transform images from lab 1 to images of lab 2 or lab 3. We will then discretize the color space accordingly. Subsequently, we will use GCNNs along with a predefined set of transforms to develop a stain invariant solution for histopathology tissue classification.
Data
In this project we plan to use public, multi-institutional datasets such as Camelyon17 (data from 5 institutions, 10 slides per institution), MIDOG (data from 4 institutions, 50 slides per institution). We will start the experiments using Camelyon17. We will train the model only on data from a single institution, we will use two centers in validation and the remaining two for testing
Results
The algorithm will be made publically available as a Docker container on https://grand-challenge.org/.
Embedding
The student will be supervised by a research member of the Diagnostic Image Analysis Group and Computational Pathology Group whose research is dedicated to analyses of histopathological slides with deep learning techniques. We have a strong collaboration with pathology experts in the field of cancer grading. The student will have access to a large GPU cluster.
Requirements
- Students Artificial Intelligence, Data Science, Computer Science, Bioinformatics in the final stages of their Master education.
- You should be proficient in python programming and have a theoretical understanding of deep learning architectures.
- Basic biological / biomedical knowledge is preferred.
Information
- Project duration: 6 months
- Location: Radboud University Medical Center
- More information can be obtained from Khrystyna Faryna (mailto: khrystyna.faryna@radboudumc.nl)
