Pretrained deep-learning IMPT planning transferred across centres with comparable plan quality

KEY POINTS

• A deep-learning IMPT model previously trained on 60 oropharyngeal cancer patients was transferred to an external proton centre. Five external cases were used for configuration and 10 independent patients for blinded evaluation.
• Patients received definitive proton therapy with 70 Gy(RBE) to the therapeutic CTV and 54.25 Gy(RBE) to bilateral elective nodal volumes. The external centre used six-beam robustly optimized IMPT with 3-mm setup and 3% density uncertainties.
• Adapting the pretrained workflow required 11 configuration iterations using five patients and was completed within two working days by a team experienced with deep-learning planning.
• In multidisciplinary blinded review, manual and deep-learning optimization plans were each preferred in 4/10 cases, while 2/10 were considered equivalent. Exactly 6/10 plans in each group were clinically acceptable without modification.
• Deep-learning plans reduced mean parotid dose from 20.1 to 17.8 Gy(RBE) (p<0.001) and oral cavity dose from 31.5 to 30.2 Gy(RBE) (p=0.002). Estimated grade ≥2 xerostomia decreased from 38.1% to 37.1%, and grade ≥3 xerostomia from 10.3% to 9.9%.
• Manual plans met all predefined clinical goals more often (8/10 vs 6/10) and had marginally higher modeled TCP. After normalization of the deep-learning plans for comparable robust target coverage, all clinical goals were met and the NTCP differences disappeared.
• Generalizability remains uncertain because evaluation involved only 10 patients at one external centre, configuration was performed by an experienced team, and plan assessment occurred in a simulated rather than prospective clinical workflow.

CLINICAL TAKEAWAY

SOURCE

Radiotherapy and Oncology