Background
Big volumes of diverse data from distributed sources are continuously produced. Health-care data is expected to be over 2000 exabytes in 2020. Supervised deep learning models, require large amount of annotated data. Considering the available data, such annotations are impractical. The ExaMode project is a collaboration between universities and industries and addresses this impracticality by applying weakly supervised deep learning based on diverse data from distributed sources.
Aim
The project aims to solve: "weakly supervised knowledge discovery of exascale heterogeneous data.’’
Radboudumc is responsible for the following objectives:
- Develop deep learning methods for detection and semantic segmentation of regions of interest in histopathology whole-slide images.
- Develop deep learning methods for whole-slide image classification using image-level target labels extracted from pathology reports.
- Develop methods of whole-slide image compression to allow end-to-end training of deep neural networks with whole-slide image data.
- Develop efficient parallel implementations of researched ANNs dealing with both shared and distributed memory environments.
- Make developed methods available via a web-based platform.
- Facilitate integration of developed methods in digital pathology workflow and development of decision support algorithms.
