Click here to

Session: Machine Intelligence Efficacy and Quality II [Return to Session]

Patient-Specific Transfer Learning to Enhance the Performance of Deep Learning Auto-Segmentation in 0.35 T MRgRT for Prostate Cancer

M Kawula1*, I Hadi1, L Nierer1, M Vagni2, D Cusumano2, L Boldrini2, L Placidi2, S Corradini1, C Belka1,3, G Landry1, C Kurz1, (1) Department of Radiation Oncology, University Hospital, LMU Munich, Germany, (2) Fondazione Policlinico Universitario Agostino Gemelli, Rome, Italy, (3) German Cancer Consortium (DKTK), Munich, Germany


MO-B-BRC-4 (Monday, 7/11/2022) 8:30 AM - 9:30 AM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Adaptive magnetic resonance-guided radiation therapy (MRgRT) requires time-consuming daily delineation and could thus benefit from auto-segmentation. This study aimed to leverage potentially available expert planning image delineations to improve auto-segmentation of CTV, bladder, and rectum on fraction images at a 0.35T MR-Linac for prostate cancer patients.

Methods: 73 planning MR images (MRIs) from a collaborating institution (cohort 1, C1) and 19 patients (19 planning, 240 fraction MRIs) from our facility (cohort 2, C2) with expert delineations were included. C1 data was used to train, validate, and test 3D-UNets for CTV, bladder, and rectum segmentation (baseline models, BMs). The BMs were subsequently fine-tuned during C2-facility-specific (FS) and patient-specific (PS) transfer learning. The FS learning used 10 C2 planning images for network parameter fine-tuning, yielding models accounting for potentially different segmentation styles at C2. The PS training fine-tuned the BMs separately for each patient using the respective planning MRI. The same 10 patients were used to optimize hyperparameters. Both the FS and the PS models shared the same test set of C2 fraction MRIs from 9 patients unseen during training. The evaluation included dice similarity coefficient (DSC), the average (HDavg), and the 95th percentile (HD95) Hausdorff distance.

Results: Compared to the BMs the PS training improved the average DSC/HD95/HDavg [mm] from 0.69/10/3.7 to 0.86/4.2/1.6 for the CTV, from 0.91/6.0/1.8 to 0.93/3.5/1.2 for the bladder, and from 0.87/5.2/1.5 to 0.90/3.7/1.1 for the rectum. FS training led to negligible improvements of BMs.

Conclusion: The PS models successfully incorporated information from expert planning delineations available in MRgRT. Main improvements were found for the CTV by delineating the correct amount of seminal vesicles and neighboring normal tissue. Moreover, PS learning corrected pronounced BM errors in the bladder and correctly determined rectum ends. Therefore, PS models have the potential to improve auto-segmentation for MRgRT.

Funding Support, Disclosures, and Conflict of Interest: The Radiation Oncology Department of the LMU Munich University Hospital has a research agreement with ViewRay. ViewRay did not fund this study, was not involved, and had no influence on the study design, data collection or analysis, and manuscript writing. This work was funded by the Wilhelm Sander-Stiftung (2019.162.1).


Segmentation, Prostate Therapy, MRI


IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email