KEY POINTS
- This observational matched cohort study from University Medical Center Groningen, the Netherlands, compared clinical outcomes after deep learning-based versus manually generated intensity-modulated proton therapy (IMPT) for oropharyngeal cancer. The original clinical populations comprised 85 patients treated with deep learning-assisted plans and 114 treated with manual plans, from which 74 matched pairs (148 patients) were selected. Matching accounted for T stage, N stage, concurrent systemic treatment, and accelerated fractionation. All patients received curative-intent treatment, prescribing 70 Gy(RBE) to the primary tumor and involved nodes and 54.25 Gy(RBE) to bilateral elective nodal regions.
- The AI workflow combined automated dose prediction with substantial expert refinement. A 3D U-Net trained on 60 patients generated predicted dose distributions, which were converted into deliverable plans through approximately 30 minutes of robust dose-mimicking optimization. Planning incorporated 21 robustness scenarios with 3-mm setup and 3% range uncertainty, followed by approximately two hours of manual refinement by an experienced planner. Conventional manual planning typically required one to two days, although this study did not prospectively quantify comparable end-to-end planning-time savings.
- No statistically significant differences in oncological outcomes were observed between planning strategies. Overall survival, progression-free survival, locoregional control, and distant metastasis-free survival were evaluated using Kaplan–Meier and Cox regression analyses. None showed a significant association with planning technique in either unadjusted models or models accounting for smoking exposure. However, the outcome evaluation was limited to relatively early follow-up, with few deaths and recurrence events, so the results do not establish long-term clinical equivalence.
- Acute and early late radiation toxicity were also comparable across 19 evaluated endpoints. Physician-rated CTCAE/RTOG and patient-reported EORTC outcomes were collected at 11 time points, from baseline through 12 months after treatment initiation. In adjusted analyses, grade 2–4 dysphagia had an OR of 0.95 (95% CI 0.67–1.48) for deep learning-based versus manual planning, while moderate-to-severe xerostomia had an OR of 0.75 (95% CI 0.48–1.18) and grade 3–4 mucositis an OR of 1.21 (95% CI 0.78–1.90). No toxicity endpoint reached the multiplicity-adjusted significance threshold of p<0.003.
- Normal-tissue dose distributions showed only small differences between the two planning approaches. Mean dose differences across nine organs of interest averaged approximately 1.6 Gy(RBE), with individual organ differences ranging from 0 to 4.1 Gy(RBE). No comparison met the Bonferroni-adjusted significance threshold of p<0.006. Organs evaluated included the parotid and submandibular glands, oral cavity, and pharyngeal constrictor muscles, suggesting that the final clinically approved plans achieved broadly similar dosimetric objectives.
- Patient matching improved comparability but did not eliminate potential confounding. Following matching, mean age was 66.2 versus 66.3 years, while concurrent systemic therapy was administered to 44.6% of patients in both groups. However, smoking exposure remained imbalanced, with mean pack-years of 24.5 in the manual cohort versus 20.4 in the AI cohort, corresponding to a standardized mean difference of 0.425. Four matched pairs were also affected by selective manual planning for technically challenging cases. Excluding these pairs in sensitivity analyses did not materially change the principal findings.
- The findings apply to a supervised AI planning workflow, not fully autonomous treatment planning. Experienced planners reviewed and manually refined every AI-generated plan before clinical delivery, and some patients were deliberately treated with manually generated plans because of technical or logistical considerations. The study was nonrandomized, involved one center, and had limited power to detect uncommon toxicity or late oncological differences. Longer follow-up and multicenter evaluation are necessary before concluding that AI-generated proton plans achieve equivalent clinical outcomes across broader patient populations.
CLINICAL TAKEAWAY
Deep learning-assisted IMPT planning with expert refinement produced early oncological outcomes and acute toxicity profiles that were not statistically different from manually generated proton plans in this matched oropharyngeal cancer cohort. These findings support the feasibility of integrating AI into supervised clinical proton-planning workflows, but do not establish equivalence, long-term safety, or the reliability of fully automated planning. Continued expert review and independent quality assurance remain essential.