Advancing offline magnetic resonance-guided prostate radiotherapy through dedicated imaging and deep learning-based automatic contouring of targets and neurovascular structures

Background and Purpose

Erectile dysfunction (ED) affects quality of life following radiotherapy for prostate cancer. Magnetic resonance imaging (MRI) planning provides superior visualization of potency-related anatomical structures compared to computed tomography (CT), enabling improved sparing. However, contouring these structures in clinical practice is time-intensive and requires expertise. Deep learning (DL) auto-contouring with MRI simulation could make neurovascular-sparing radiotherapy more accessible.

Material and Methods

High-resolution 3D T2-weighted SPACE MRI sequences (<1 mm³ resolution) were obtained for 50 patients in treatment position. An expert uro-radiation oncologist contoured erectile function-related anatomy (neurovascular bundles [NVB], pudendal arteries [IPA], penile bulb [PB], corpora cavernosa [CC]) and target structures (prostate [PR], seminal vesicles [SV]). Forty datasets trained and ten tested a 3D nnU-net model. DL-generated contours were geometrically evaluated (surface Dice Score [sDSC], mean surface distance [MSD]) and validated by blinded expert review.

Results

DL auto-segmentation achieved an average sDSC of 0.82 (IPA: 0.93, NVB: 0.71, PB: 0.84, CC: 0.90, PR: 0.74; SV: 0.79) and average MSD of 0.74 mm (IPA: 0.61 mm; NVB: 0.88 mm; PB: 0.63 mm; CC: 0.47 mm; PR 0.83 mm; SV: 1.01 mm). Blinded ratings showed no significant differences between DL and expert contours, except for pudendal arteries (Mean DL vs. expert; NVB: 82 vs. 85; PB: 86 vs. 88; CC: 83 vs. 88; PR 81 vs. 83; SV 78 vs. 81 all p > 0.05; IPA: 82 vs. 89; p = 0.028).

Conclusion

Combining high-resolution MRI simulation with DL postprocessing enables accurate auto-contouring for MR-guided SBRT planning, potentially advancing neurovascular-sparing radiotherapy beyond current standards.