Auto-segmentation of organs at risk in prostate cancer patients for MR-Linac

Abstract

Accurate and efficient delineation of pelvic organs-at-risk is essential for MRI-based prostate radiotherapy. Manual contouring is time-consuming and prone to observer variation. This study evaluates whether deep-learning auto-segmentation can provide reliable contours for prostate MRI planning by comparing a slice-wise Swin-UNet with a volumetric nnFormer. An expertly annotated, anonymised T2-weighted MRI cohort was used for development and testing (training n = 261; testing n = 66). Model performance was assessed against expert contours using the Dice Similarity Coefficient (DSC) and the 95th percentile Hausdorff distance, alongside a qualitative visual review to characterise typical failure modes. On the test set, nnFormer outperformed Swin-UNet for the bladder with DSC 0.77 versus 0.66 and HD95 1.69 versus 4.77 mm, and for the clinical target volume (CTV) with DSC 0.74 versus 0.63 and HD95 1.64 versus 3.12 mm. For the rectum, nnFormer achieved a lower HD95 of 1.93 versus 4.39 mm but a slightly lower DSC of 0.65 versus 0.75. Qualitatively, nnFormer produced more spatially coherent and anatomically faithful contours with narrower case-to-case variability. In contrast, Swin-UNet was more susceptible to artefacts and slice-to-slice intensity variation, with common failure modes including under-segmentation of overfilled bladders and fragmented rectal walls in the presence of gas. These findings indicate that automated segmentation of the clinical target volume and organs-at-risk on prostate MRI is feasible with both architectures. However, the volumetric nnFormer provides superior boundary fidelity and overall accuracy for bladder and CTV. Further research is required to confirm and expand these results. This includes validating the approach on MR-linac systems across multiple centres, incorporating uncertainty-aware quality assurance to better guide human review, and prospectively assessing how the method affects editing time and dosimetric outcomes in MRI-only and MR-guided adaptive workflows.