“My ideas grow better when transplanted into another mind than in the one where they sprang up. “ Oliver Wendell Holmes, Sr.

Hi! I am ESR2, Yuandou Wang, and I will share some experiences with the secondment at UiS.


At the end of January 2022, I flew to Stavanger, and began my secondment at UiS in the beginning of February. As planned in the secondment agreement, during the three months, I will gain knowledge about the Whole-Slide-Image (WSI) preprocessing, anonymization and data privacy techniques, and artificial intelligence applied for HR-NMIBC characterization in BMDLab, UiS. In addition, I will explore workflow technologies, especially for distributed medical image preprocessing and AI pipelines in the medical research domain, by cooperating with biomedical image analysis experts – ESR4 (Neel Kanwal) and ESR5 (Saul Fuster Navarro), and Prof. Kjersti Engan.

In the first month of the secondment, I briefly introduced myself in the BMDLab and shared with our research work in the group at UvA. Then, ESR4 and ESR5 gave lectures on WSI preprocessing and classification of histological images to introduce their research work, respectively. I learnt a lot from their nice presentations and discussions regarding the computational performance bottlenecks. They are facing with challenges in the AI model training process, for example, 1) limited GPU resources with heavy computing tasks, 2) limits for the parallelism of some critical steps strongly influence the total completion time of the AI pipelines.

To cope with the above issues, one of solutions I am thinking about is to make such AI pipelines as manageable workflows. The start point of this work is to extract the structure of the AI pipelines. Through actively discussing with ESR4 and ESR5 about the code structure, we are going to further highlight and define the atomic steps, and dependencies among the atomic steps in program. My task currently is to represent the atomic steps as functionality tasks, clarify the dependencies among these tasks, and extract the whole AI training process as the common workflow language (CWL). On this basis, we can make the AI pipelines as visual and executable workflow, and further design scheduling algorithms to create execution plan for optimizing the workflow execution, such as allocating dedicated cloud infrastructure to speed up the total workflow completion time.

The active discussion with AI experts in BMDLab continues, and I am enjoying the bounce around our ideas with a group of experts. Enjoy beautiful life in Stavanger, Norway!

Discussion more details with ESR4 about the digital diagnosis workflows.


Whiteboard & Handwriting: Brainstorming with Prof. Kjersti Engan

Yuandou Wang – ESR2