Work Package 4

Management Team

Main Objectives 

WP4 will develop tools for accessing, uploading, annotating and mining digital slides in a two-tiered approach. First, tools to provide access to WSI will be built upon Cytomine, an established open-source, cross-platform framework, to ease access to WSI, including annotations and visualisation of algorithm results. Open- source tools will be developed to facilitate efficient collection, dicomisation and pseudonymisation/anonymisation of WSI and meta-data. Secondly, WP4 will develop a strategy for the development of new and generic AI-models and algorithms in order to facilitate deep-learning and mining of WSI data.

Main Tasks

Achieving these objectives, WP4 will be done by:

a) Developing task-agnostic models: WP4 partners aim to overcome the unending task-specific “collect data – train model – test model” cycle by focusing on generic, task agnostic interpretable AI models based on out-of-distribution data analysis and deep transfer learning.

b) Developing new techniques for extraction of signatures based on imaging data: by using new techniques to directly predict slide level metadata through the extraction of signatures from imaging data, we reduce the need for detailed annotations and allow model development on large-scale datasets. In addition, through combination with our efforts in interpretable AI models, we will be able to identify morphological features or biomarker substrates relevant for certain slide-level labels. This will also form the basis for development of content-based image retrieval techniques, specifically tailored for histopathological images.

c) Using federated learning strategies: we will develop a federated learning strategy, building DL models with data from multiple sources without the requirement for the data to reside in a central location (e.g. using private data that cannot leave the data source for legal or ethical reasons). This approach will allow us to develop DL models to analyse, for example, rare diseases which (by legal definition) cannot be anonymised.

d) Setting up an integrated app store: Through setting up an integrated app store for DL models we will provide algorithm developers with tools to a) reach a large audience, b) evaluate their models in real life setting on unprecedented numbers of slides and c) provide a potential avenue for monetisation.