Deep learning-enhanced image registration for accelerating daily adaptive magnetic resonance imaging-guided prostate radiotherapy.
Authors
Affiliations (1)
Affiliations (1)
- Department of Radiotherapy, UMC Utrecht, Heidelberglaan 100, 3508 GA, The Netherlands.
Abstract
Daily auto-contouring remains a workflow bottleneck in magnetic resonance-guided adaptive prostate radiotherapy (MRgRT). This study proposes and clinically validates a novel deep learning-enhanced deformable image registration (DIR) solution to accelerate this critical step. A hybrid framework combining a 3D nnU-Net segmenting bladder/rectum on planning/daily MRI with an in-house DIR algorithm was implemented for 5-fraction prostate MRgRT ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>5</mn> <mo>×</mo> <mn>7.25</mn></mrow> </math> Gy) on an MR-Linac. The DIR uses nnU-Net contours to propagate target and organs-of-interest structures. The solution was clinically deployed and evaluated in 275 patients/1375 fractions. Evaluation following clinical introduction, has shown a median contouring time of <math xmlns="http://www.w3.org/1998/Math/MathML"><mo>≈</mo></math> 190 s, halving the time required by the previously-employed vendor-provided solution. Quantitative evaluation showed high agreement with clinically approved contours: Dice similarity coefficients <math xmlns="http://www.w3.org/1998/Math/MathML"><mo>></mo></math> 0.9 and 95th percentile Hausdorff distances <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 2.0 mm for most structures. The implemented solution demonstrated reliable, high-accuracy daily auto-contouring, significantly accelerating MRgRT workflows. It has become our institutional standard for prostate treatments. Future work will extend this approach to additional treatment sites and modalities.