BLUE RIBBON POSTER MULTI-DISCIPLINARY: Mspock: Multi Contrast Sparse-to-Precise Organs-at-Risk Contouring with Prior Knowledge for Pancreatic MR-Linac Radiotherapy
Abstract
Purpose
To develop an MR organs-at-risk (OAR) auto-contouring pipeline that is robust in severely limited-data settings across contrasts, specifically for MR-Linac pancreatic cancer treatments where full-volume manual contouring can take ~8h. We developed mSPOCK (multi-contrast Sparse-to-Precise Organs-at-risk Contouring with prior Knowledge), which takes sparse contours and generates precise full contours using a conditioned nnU-Net.
Methods
199 MR images (validation/test:32/12) with different contrasts and nine OARs (small bowel/duodenum/left kidney/right kidney/large bowel/liver/spinal canal/spleen/stomach) were used: 42 Unity MR-Linac cases (manually contoured; approved by two clinicians) and 157 abdominal MR scans screened from TotalSegmentatorMR dataset. Sparse inputs (a per-OAR minimum set of rough 2D contours on few slices) were simulated using structure-guided deformation augmentation method (sgDefAug) with slice removal and dilation/erosion. Then, missing slices were added by morphology-based interpolation to generate OAR-specific deformations. A conditioned nnU-Net was trained to predict full contours from paired images and simulated contours. For testing, two oncology consultants provided sparse contours per OAR. Performance was assessed using Dice similarity coefficient (DSC), average surface distance (ASD), and 95th percentile Hausdorff distance (HD95), and compared with TotalSegmentatorMR. Outputs were edited to clinical acceptability and timed. Dosimetric impact was evaluated via D0.1cm³ and D50% differences versus ground-truth relative to the prescribed dose of 40Gy to the planning target volume.
Results
Training required 57h; inference was <30s. mSPOCK achieved high accuracy and outperformed TotalSegmentatorMR (DSC 0.92±0.03, ASD 1.48±0.94mm, HD95 6.44±4.96mm vs 0.76±0.10, 4.92±2.70mm, 15.16±8.70mm). Clinicians spent ~11 min on sparse contouring and ~22 min editing per case to clinical acceptability, versus hours for full-volume manual contouring. Plan evaluation showed small relative dose differences versus ground-truth (D0.1cm³: 0.7±6.4%, D50%: 0.2±0.8%), supporting the feasibility of faster contour generation with minimal dosimetric difference.
Conclusion
By converting a handful of sparse slices into precise OARs within minutes, even with limited data, mSPOCK shortens contouring workflows and helps reduce treatment delays.