Copenhagen, Denmark

ESTRO 2022

Session Item

May 08
10:30 - 11:30
Room D4
MR-guided radiotherapy
Marcel van Herk, United Kingdom;
Vivian van Pelt, The Netherlands
Proffered Papers
11:10 - 11:20
Patient specific deep learning contour propagation on prostate magnetic resonance linac patients
Samuel Fransson, Sweden


Patient specific deep learning contour propagation on prostate magnetic resonance linac patients

Samuel Fransson1, David Tilly1, Robin Strand2

1Uppsala University Hospital, Medical Physics, Uppsala, Sweden; 2Uppsala University, Department of Information Technology, Uppsala, Sweden

Show Affiliations
Purpose or Objective

Devices combining MR-scanners and Linacs for radiotherapy, called MR-Linacs, requires contouring on a daily basis to be used to its fullest. Currently, deformable image registration (DIR) algorithms propagate contours from reference scans, however large shape and size changes can be troublesome. Artificial neural network (ANN) based contouring models not relying on DIR algorithms alleviate this issue. However, the requirement of similarity of the training and inference dataset poses an issue with potential highly variable contrast of MR-images, along with patient specific target definition not present in training dataset. To alleviate this problem of scarcity of data, we propose patient specific networks, trained on a single dataset for each patient, for contouring onto the following datasets in a adaptive MR-Linac workflow. 

Material and Methods

MR-scans from eight prostate patients treated on the MR-Linac (6.1 Gy x 7 fx) at our institution along with contours of Clinical Target Volume (CTV), bladder and rectum were utilized. U-net shaped models were trained based on the image from the first fraction of each patient, and subsequently applied onto the following treatment images. Results were compared with manual contours in terms of the DICE overlap as well as Added Path Length (APL) which correlates with recontouring time. As benchmark, contours propagated through the clinical DIR algorithm were similarly evaluated.


In terms of DICE overlap the ANN output was 0.91±0.03, 0.94±0.04 and 0.82±0.09 while for DIR 0.92±0.03, 0.92±0.08, 0.86±0.05 for the CTV, bladder and rectum respectively. Similarly, APL results where 3479±1835, 7356±4391 and 6832±2362 for ANN and 2853±1687, 8571±4654 and 6395±2663 voxels for DIR.


Patient specific artificial neural network models trained on a single dataset are feasible with comparable accuracy to DIR.