Conditional organs-of-interest segmentation with plausible inter-fraction variation simulation for pancreatic...

Manual contouring of organs of interest (OoIs) is a major bottleneck in pancreatic magnetic resonance-guided online adaptive radiotherapy (oART). We developed C-SegDeform, a data-efficient conditional segmentation framework that used structure-guided deformation-based augmentations to simulate plausible inter-fraction anatomical variation and leveraged organ-specific conditioning as an alternative to registration-based contour propagation (Prop-ROIs) in limited-data settings.

Forty balanced 3DVane images from 12 patients were manually contoured and pre-processed, including duodenum, both kidneys, liver, large and small bowel, spinal canal, spleen, and stomach. The training dataset (26 images) was augmented by simulating plausible session images via structure-guided deformations. Data was arranged for conditional segmentation and leave-one-out cross-validation using the nnU-Net framework. Analysis included geometric (DSC, average surface distance (ASD), and 95th percentile of Hausdorff distance (HD₉₅); against Prop-ROIs and TotalSegmentator MRI), dose (via 8 treatment plans, measuring $D_{0.1c{m}^3}$ and $D_{50\text{\%}}$ discrepancies relative to prescribed dose), and clinical (Likert scale and post-auto-contour editing times by two oncology consultants) assessments.

C-SegDeform (DSC: $0.88\pm 0.10$ , ASD: $2.9\pm 2.4\mathrm{mm}$ , HD₉₅: $10.5\pm 8.9\mathrm{mm}$ ) outperformed Prop-ROIs ( $0.70\pm 0.17$ , $8.5\pm 9.9\mathrm{mm}$ , $16.7\pm 9.8\mathrm{mm}$ ) and TotalSegmentator ( $0.75\pm 0.16$ , $5.1\pm 3.4\mathrm{mm}$ , $14.3\pm 10.6\mathrm{mm}$ ), and showed smaller relative dose deviations from ground-truth contours ( $D_{0.1c{m}^3}$ : $2.2\text{\%}\pm 2.5\text{\%}$ , $D_{50\text{\%}}$ : $0.5\text{\%}\pm 0.9\text{\%}$ ) than Prop-ROIs ( $6.8\text{\%}\pm 5.7\text{\%}$ , $3.3\text{\%}\pm 4.0\text{\%}$ ). Clinicians ranked 98% of C-SegDeform contours as requiring no or minor edits (vs. 70% for Prop-ROIs), with editing times under 6 min and full OoI generation completed in under 40 s.

C-SegDeform provided a robust alternative to contour propagation, generating accurate contours rapidly with minimal edits, thereby reducing on-couch time and streamlining the workflow.