Head-And-Neck IMRT Auto-Planning Through Fluence Map Prediction Using Progressive Growing of Generative Adversarial Networks
X Li1*, QJ Wu1, Q Wu1, C Wang1, Y Sheng1, W Wang1, H Stephens1, F Yin1, Y Ge2, (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC
Presentations
TH-F-TRACK 4-4 (Thursday, 7/29/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Purpose: To develop an IMRT auto-planning algorithm for head-and-neck cases, where optimal fluence maps were predicted from patient anatomy using Progressive Growing of Generative Adversarial Networks (PGG).
Methods: PGG enhances the traditional architecture of Generative Adversarial Networks (GAN) with a multi-resolution design that is potentially more effective at capturing the complexity of the head-and-neck cases. In PGG, input was a series of 128x128 2D projections representing patient anatomy for 9 template beam angles. The networks were trained in 6 stages, where the output fluence map dimensions were 4x4, 8x8,…, and 128x128 for 9 beams. Upon a new stage, up-sampling layers and convolutional layers were added to the generator. Generator output gradually changed from up-sampling of previous stage’s output to the average of previous output and new convolutional layers’ result. Ground truth (128x128) was downsampled by average filters to the same dimension with the output for loss calculation. Furthermore, an additional discriminator taking higher resolution input was added to the existing discriminator group with a gradually increasing weight. The discriminator group contains discriminators using (downsampled) ground truth, anatomic input, and generator prediction with different resolutions. Training, validation, and test patient numbers were 200, 16, and 15. For test patients, last stage’s output was post-processed into predicted fluence maps and was converted into IMRT plans in TPS (PGG plans).
Results: Preliminary results showed that PGG plans had overall acceptable plan quality for test patients, except that the maximum dose is high (whole-body D2cc=123.0%). PGG plan D_mean in parotid L/R(22.2±2.5Gy/23.4±3.0Gy), oral cavity(24.6±7.0Gy), larynx(22.1±5.7Gy), pharynx(35.4±2.9Gy) showed no clinically relevant difference from benchmark plans(23.1±2.0Gy/23.9±2.3Gy/23.9±4.3Gy/22.7±4.8Gy/34.7±2.5Gy). PGG plan D_0.1cc in brain stem(17.5±3.8Gy) and cord+5mm(30.5±2.5Gy) was higher than in benchmark plans(15.5±2.7Gy/25.8±1.9Gy).
Conclusion: The PGG-based head-and-neck IMRT planning algorithm has demonstrated encouraging plan quality. Further development and training are underway to improve performance.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH grant (#R01CA201212) and Varian master research agreement.
Handouts
Keywords
Modeling, Treatment Planning, Intensity Modulation
Taxonomy
Contact Email