Author: Mark Bowers, Gabriel Carrizo, Jimmy Caudell, Vladimir Feygelman, Kevin Greco, Christian Hahn, Jihye Koo, Kujtim Latifi, Fredrik Lofman, Jacopo Parvizi, Muqeem Qayyum, Caleb Sawyer 👨🔬
Affiliation: RaySearch Laboratories, Moffitt Cancer Center 🌍
Purpose: Head and neck (H&N) radiotherapy planning is complex, with multiple competing objectives. We endeavored to improve efficiency of planning by developing a deep learning (DL) model trained to predict an ideal patient specific dose distribution and optimize and output a deliverable plan by mimicking the predicted distribution.
Methods: A 3D U-net architecture combines an encoder and decoder to generate three-dimensional dose distributions. The model processes patient-specific target volumes and organs at risk as binary input maps. These representations enable the network to focus on geometric and spatial relationships without detailed CT density information. The architecture implements 3D operations, including convolutions, maxpooling, and transpose-convolutional layers. The resulting dose distribution aids optimization and guides manual planning. The model was trained on 54 oropharyngeal cancer patients treated with volumetric modulated arc therapy with two planning target volumes (PTVs) (70 Gy to primary tumor, 56 Gy to elective region). The DL model was then validated on 10 other patients with previously delivered manual plans (MP).
Results: Mean doses (Gy) and standard deviation for the plans produced by DL and MP to PTV70 D95 were 70.95±0.31 and 70.26±0.16, PTV56 D95 56.97±0.24 and 56.39±0.20, gross tumor volume (GTV) D100 70.56±0.43 and 70.04±0.38. The DL plans were produced in under 30 minutes while dosimetrists reached the MP goals after several hours. Paired t-tests showed MP had significantly lower D0.03 and D95, while higher D100 across PTVs and GTV. Mean esophagus, oral cavity, parotids, submandibular glands, constrictors, and cord max doses were not significantly different.
Conclusion: DL plans were clinically comparable to their manually planned equivalents. As DL plans can be further optimized, this may reduce overall planning time, increasing efficiency of H&N treatment planning. This may improve plan quality, particularly in low volume clinics, and availability of adaptive planning.