MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Lightweight Deep-Learning Model for Automatic IMRT Planning Via Fluence Map Prediction with a 2.5D Implementation: A Study of Head-And-Neck IMRT Application

C Wang1*, X Li1, Y Sheng1, J Zhang1, K Lafata1, F Yin1, Q Wu1, Y Ge2, Q Wu1, (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC

Presentations

(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 3

Purpose:
To develop a deep-learning model for fully-automated head-and-neck IMRT plan generation with a novel lightweight design.


Methods:
This model generates radiation fluence maps of a 9-beam template without time-consuming inverse optimization. The model implements a novel deep-learning network, PyraNet, which connects 14 classic ResNet blocks in pyramid-like concatenations. For a voxel-of-interest in a patient’s 3D volume, the model generates a series of customized HU/contour 2D projection patches (in 21-by-21 pixel size) at 9 template beam angles. These patches are stacked as PyraNet inputs(2.5D presentation), and PyraNet predicts the radiation intensities of 9 beamlets in 9 beams that intersect at the voxel-of-interest. By iterating through the 3D volume, all beamlets’ predictions are synthesized into 9 radiation fluence maps, which are automatically imported into a commercial TPS for plan finalization.

The model was built upon 231 oropharyngeal plans in an anonymized IMRT plan library, where 200/16/15 plans were assigned for training/validation/test, respectively. Only the plans involving primary PTV were selected and were normalized to 44Gy(2Gy/fx). The model training was regularized by the mean-absolute-error of the synthesized radiation fluence maps against the ground-truths. For model evaluation, 15 plans(DL-plans) generated by the model were compared with the ground-truth test plans. Key dosimetric metrics were compared by Wilcoxon Signed-Rank tests.


Results:
All 15 DL-plans met institutional planning guidelines. After PTV coverage normalization, mean dose of left-parotid(23.4±3.2Gy), right-parotid(23.7±3.8Gy) and oral cavity(25.0±5.4Gy) in DL-plans were comparable to the library plan results (23.1±2.0Gy/23.9±2.3Gy/23.9±4.3Gy) without clinically-relevant differences. Max dose results D0.1cc of brainstem(14.7±2.2Gy) and cord+5mm(26.3±1.6Gy) in DL-plans were comparable similar to library plan results (15.5±2.7Gy/25.8±1.9Gy), but body max dose results were slightly higher in DL-plans.


Conclusion:
The developed deep-learning model can generate a satisfying head-and-neck IMRT plan with a fully-automated execution. With its lightweight design, the model is suitable for further validations and applications in the clinical environment.

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

Keywords

Treatment Planning, Intensity Modulation, Modeling

Taxonomy

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

Contact Email