MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

An Artificial Intelligence-Driven Agent for Rapid Head-And-Neck IMRT Plan Generation Using Conditional Generative Adversarial Networks (cGAN)

X Li1*, Y Sheng1, J Zhang1, W Wang1, F Yin1, Q Wu1, Y Ge2, Q Wu1, C Wang1, (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC

Presentations

(Monday, 7/13/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose:
To develop an Artificial-Intelligence(AI)-driven agent for fully-automated rapid head-and-neck IMRT plan generation without time-consuming dose-volume-based inverse planning.

Methods:
This AI agent implements a conditional Generative Adversarial Networks(cGAN) architecture. The generator, PyraNet, is a novel deep-learning network that implements 28 classic ResNet blocks in pyramid-like concatenations. The discriminator is a customized 4-layer DenseNet. The AI agent generates multiple customized 2D projections at 9 template beam angles from a patient’s 3D CT volume and structures. These projections are stacked as 4D inputs of cGAN, in which 9 radiation fluence maps of template beam angles are generated simultaneously. The predicted fluence maps are imported into TPS automatically for plan finalization.
The AI agent was built and tested upon 231 oropharyngeal plans in an anonymized head-and-neck IMRT case library. Only the primary plans in the sequential boost regime were selected. All plans were normalized to 44Gy prescription(2Gy/fx). 200/16/15 plans were assigned for training/validation/independent test, respectively. A customized Harr wavelet loss was adopted for fluence map comparison during training. For test cases, isodose distributions in AI plans and library plans were qualitatively evaluated. Key dosimetric metrics were compared by Wilcoxson Signed Rank tests.

Results:
AI plans met all institutional planning guidelines. Isodose gradients outside of PTV in AI plans were comparable with library plans. After PTV coverage normalization, D(mean) of left-parotid(23.1±2.4Gy), right-parotid(23.8±3.0Gy) and oral cavity(24.7±6.0Gy) were comparable to the library plan results(23.1±2.0Gy/23.9±2.3Gy/23.9±4.3Gy) without clinical differences. AI plans achieved comparable results for brainstem and cord+5mm D(0.1cc), but body D(max) results were higher than the library plan results. The AI agent needs only 2s/case for radiation fluence map prediction.

Conclusions:
With fully-automated execution, the developed AI agent can generate complex head-and-neck IMRT plans with satisfying dosimetry quality. With rapid and fully-automated implementation, it holds great potential for clinical applications in pre-planning decision-making and real-time planning.

Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by NIH R01CA201212

Keywords

Treatment Planning, Radiation Therapy, Modeling

Taxonomy

TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email