KEY POINTS
- The dose-prediction model was trained on 266 adaptive MRI datasets from 25 prostate cancer patients treated with 60 Gy in 20 fractions using nine-beam IMRT on a 1.5-T MR-linac. Training included 216 fractions from 17 patients, with 50 fractions from eight unseen patients used for validation.
- AutoAdapt used a physics-aware Swin UNETR model to predict beam-specific dose distributions, convert them into optimization constraints, and feed those constraints into a commercial treatment-planning system. Clinical testing used the reference MRI and first adaptive fraction from 10 additional patients.
- Reference-plan constraints produced clinically acceptable results without adaptation in only four of ten cases, including two cases with target underdosage exceeding 1 Gy. AutoAdapt generated plans judged clinically acceptable without further adjustment in all ten cases, while manual planning met every formal objective in all ten.
- AutoAdapt met every prespecified objective in seven of ten patients. Two plans achieved a PTV60 D50% of 59.9 Gy rather than >60.0 Gy, and one exceeded the bladder V60Gy limit by 0.1 cm³; all three deviations were judged marginal.
- AutoAdapt reduced median rectal D0.035cm³ from 59.7 to 59.1 Gy compared with manual planning (p=0.01). Manual planning produced lower rectal V20Gy—44.0% versus 50.9% (p=0.02)—while target and bladder metrics did not differ significantly.
- The pipeline required a median of approximately 29 seconds more than manual planning, but model inference took only 3.2 seconds and direct planner input was under 20 seconds. Most additional time came from data export, preprocessing, and transferring constraints into the closed commercial system.
- AutoAdapt plans were less complex, with mean plan-aperture modulation of 0.276 versus 0.322 (p=0.007) and 6.4% fewer monitor units (p=0.006). Generalizability remains uncertain because testing involved ten cases, one institution, one beam configuration, and one fractionation schedule.
CLINICAL TAKEAWAY
AutoAdapt could reduce active planner workload during prostate MR-guided adaptation without closing off the option for human intervention. It is a promising semi-automated planning assistant, not a validated autonomous system, and its different prioritization of rectal high- and low-dose exposure requires explicit local commissioning.