KEY POINTS
- The retrospective analysis compared 12 clinical and artificial intelligence-assisted VMAT plans from 11 patients with head and neck cancer. The cohort deliberately included difficult anatomy, including a bulky tumor and a separate replan after tumor rupture.
- The study evaluated the initial 40 Gy in 20 fractions phase of a sequential treatment course planned to a total dose of 70 Gy. Clinical plans were created manually in Eclipse, while artificial intelligence-assisted plans used RatoGuide deep-learning dose prediction to generate seven dose-derived optimization structures.
- Mean planning target volume dose was similar between approaches, but the artificial intelligence plans had slightly lower D98%: 34.00 versus 34.47 Gy, and higher D2%: 42.23 versus 41.69 Gy and Dmax: 45.53 versus 43.72 Gy; all three differences were statistically significant.
- Dose conformity was effectively unchanged. The mean Paddick conformity index was 0.479 for clinical plans and 0.478 for initial artificial intelligence-assisted plans, with p = 1.00.
- Artificial intelligence-assisted planning reduced spinal cord planning-risk-volume maximum dose from 27.50 to 23.97 Gy and mean dose from 17.69 to 12.92 Gy, both with p < 0.01. Brainstem maximum dose decreased from 23.97 to 21.59 Gy, and mean dose from 7.73 to 5.49 Gy.
- Parotid sparing was not uniformly better. Artificial intelligence-assisted plans lowered ipsilateral parotid mean dose from 21.44 to 19.28 Gy, but increased contralateral parotid mean dose from 16.82 to 20.16 Gy, reflecting differences between the model and the institution’s preference to prioritize one parotid gland.
- Three of the 12 initial artificial intelligence plans produced surface hotspots of 119.94%, 121.00%, and 123.65% of prescription. Manual re-optimization reduced these to 109.68%, 109.55%, and 111.03%, bringing all plans below the institutional 115% limit.
- Mean monitor units were comparable between approaches, while planning time decreased from approximately two hours to 30 minutes. Cases requiring manual correction needed an additional 15–30 minutes, reinforcing that the tested workflow was assisted rather than autonomous.
CLINICAL TAKEAWAY
Deep-learning dose prediction may reduce planning time and provide a more consistent starting point for complex head and neck VMAT. However, the surface hotspots and parotid trade-offs show why expert review remains essential; these small retrospective data support a human-in-the-loop workflow, not autonomous clinical planning.