Successful collaboration between LUMC and LIACS on AI for radiotherapy
Daily-adapted radiotherapy can help to more precisely target radiation dose to tumors compared to the current clinical practice, while avoiding radiosensitive organs-at-risk in the surrounding area. A main obstacle however is that new treatment plans need to be created every day, which is a manual and time-consuming process. A team from LUMC and LIACS recently created AI technology that can do this fully automatically with promising accuracy and in real-time.
The collaboration started in 2019 when Computer Science student Laurens Beljaards joined the LKEB lab at the Department of Radiology, LUMC, for his master thesis. He teamed up with PhD student Mohamed Elbially Elmahdy to explore an idea that could potentially not only reach clinically required accuracies, but also do this within the time constraints that the clinical procedure requires. They focused on the problem of auto-contouring of prostate CT scans. Such a problem can generally be tackled by either a direct segmentation of the daily scan, or by registration of a planning scan, each method with its own strengths and weaknesses. By leveraging the concept of multi-task learning, the team created a convolution neural network architecture that combined the two methods in a single framework. This allowed the two methods to share certain information, resulting in substantial benefits over either one. Promising accuracies of around 1 mm were obtained, with a processing time of less than a second.
The results were recently published in the IEEE Access journal. The principal investigators Marius Staring (LUMC) and Fons Verbeek (LIACS) are very pleased with the collaboration and look forward to further building strong connections between the two institutes.
M.S. Elmahdy, L. Beljaards, S. Yousefi, H. Sokooti, F. Verbeek, U.A. van der Heide and M. Staring, "Joint Registration and Segmentation via Multi-Task Learning for Adaptive Radiotherapy of Prostate Cancer," IEEE Access, 2021.