KEY POINTS
- This retrospective study evaluated commercial ETHOS auto-contouring in 40 postoperative oral tongue cancer patients, comparing AI contours with clinically approved physician contours for six OARs: brainstem, spinal cord, bilateral parotids, esophagus and mandible.
- Geometric agreement was high overall, with mean Dice coefficients ranging from 0.81 to 0.99. The mandible performed best (0.99) while the spinal cord performed worst (0.81), with spinal-cord HD95 reaching 3.35 mm.
- Parallel organs were comparatively robust to contour differences. Mean parotid dose differences were only +0.15 Gy ipsilaterally and −0.41 Gy contralaterally, with strong concordance between AI and physician contours.
- Small geometric errors mattered more for serial organs. Spinal-cord Dmax was 1.98 ± 2.70 Gy higher using AI contours, while brainstem Dmax differed by −1.45 ± 2.41 Gy; both differences were statistically significant.
- Re-optimization did not remove the serial-organ sensitivity. Brainstem and spinal-cord Dmax differences persisted at approximately −1.78 Gy and +1.15 Gy, respectively, despite preserved target coverage and plan quality.
- Clinical review reflected the same pattern. 90% of mandible and 78–80% of parotid contours were considered directly acceptable, compared with only 55% for spinal cord and 60% for esophagus; a small proportion of the latter were rated unacceptable.
- AI generated the six OARs in approximately 1.5 minutes. Clinician review and correction added about eight minutes, bringing the total AI-assisted workflow to 9.5 minutes versus approximately 20 minutes manually for the same structures.
CLINICAL TAKEAWAY
Commercial auto-contouring can meaningfully reduce head-and-neck planning workload, but high Dice scores should not be treated as proof of dosimetric safety. Physician review remains particularly important for narrow serial structures such as the spinal cord, where millimetric boundary differences can change Dmax by 1–2 Gy.