KEY POINTS
- The EvoDL framework was clinically implemented in 275 patients with intermediate-risk prostate cancer receiving 5 × 7.25 Gy on an Elekta Unity MR-Linac, covering 1,375 treatment fractions between November 2024 and January 2026.
- EvoDL combines a 3D nnU-Net with deformable image registration: the network segments the bladder and rectum on planning and daily T2-weighted MRI, and these contours guide the Evolution registration algorithm used to propagate the remaining planning structures. The nnU-Net had been trained on MRI and clinical contours from 115 prostate cancer patients.
- In pre-implementation testing across 105 inter- and intrafraction image pairs from 10 patients, the CTV required no adjustment in 70/105 cases and only minor adjustment in 31/105; the rectum required no adjustment in 101/105. All contour verification and correction was completed within 2 minutes.
- After clinical deployment, median end-to-end contouring time fell from approximately 470 seconds (7.8 minutes) with Monaco 6.2.1 to 190 seconds (3.1 minutes) with EvoDL (p < 0.01). The EvoDL measurement included approximately 90 seconds of computational latency plus manual verification and correction.
- Agreement between EvoDL proposals and clinically approved contours was high: median Dice coefficients were 0.98 for bladder, 0.97 for rectum, and 0.96 for CTV, with median 95th-percentile Hausdorff distances of 0.70 mm, 0.70 mm, and 2.00 mm, respectively. More than 75% of evaluations had Dice values of at least 0.95 and corresponding Hausdorff distances below 2.13 mm.
- Clinically important failures were uncommon but not absent: the deep-learning component failed to provide a reliable bladder contour in 10/1,375 fractions, while registration failed in another 9/1,375 fractions with particularly large planning-to-daily displacements exceeding 2.5 cm. All were detected and corrected by the treating RTT before treatment.
- Contouring time did not differ significantly across the five fractions (p = 0.85) and remained broadly stable during the first 3.5 months after implementation. The authors note that human contour verification alone can still add approximately 90–120 seconds, even when no editing is required.
CLINICAL TAKEAWAY
For centres performing daily adaptive prostate MRgRT, a hybrid deep learning–DIR approach can remove several minutes from one of the main online workflow bottlenecks while preserving high contour agreement. The evidence is stronger than a purely technical validation because the system was used across 1,375 clinical fractions, but it remains a single-centre experience with a historical vendor comparator and rare major failures requiring human detection. This is technically important and potentially workflow-changing, not evidence of improved clinical outcomes.