MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automatic IMRT Planning Via Static Field Fluence Prediction (AIP-SFFP): A Novel Local Attention Deep-Learning Design for Head-And-Neck IMRT Application

X Li1, J Zhang1, Y Sheng1, Y Chang1, H Stephens1, Q Wu1, F Yin1, Y Ge2, Q Wu1, C Wang1*, (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
To investigate a novel local-attention deep-learning(DL) design of Automatic-IMRT-Planning-via-Static-Field-Fluence-Prediction(AIP-SFFP) for Head-and-Neck(H&N) application.


Methods:
When applying AIP-SFFP on H&N cases, a novel local-attention DL design was employed for local information focusing and fluence map prediction quality enhancement. This novel AIP-SFFP algorithm employes a customized DL network, PyraNet, which implements 18 classic ResNet blocks in pyramid-like concatenations. This network predicts the intensity value of a single pixel in a beam’s radiation fluence map using a series of customized 2D projections. For each pixel-of-interest, the 2D projections are generated from the patient’s 3D CT volume and PTV/CTV/organs-at-risk structures with a limited 13x13 pixels field-of-view. A radiation fluence map is generated via iterative implementations of PyraNet of all pixels, and a total of 9 maps for a 9-beam arrangement template are generated by 9 independently-trained PyraNet models. The generated 9 radiation fluence maps are automatically imported into a commercial TPS as plan parameters.

This AIP-SFFP was trained using the primary plans (44Gy in 22fx) of 216 oropharyngeal sequentially-boosted cases from a research case pool. All research plans used the same 9-beam arrangement template. Mean absolute error was used as the loss function during the training of each beam’s PyraNet model. For independent tests, AIP-SFFP generated plans for another 15 cases from the research pool. The AIP-SFFP plans were compared against the research plans by dosimetric parameter evaluation.


Results:
AIP-SFFP plans achieved reasonable spatial dose distributions. D(mean) of oral cavity (25.2±6.5Gy), left-parotid(23.9±3.1Gy), right-parotid(23.9±3.8Gy), larynx(23.8±5.6Gy) and pharynx(34.7±2.5Gy) in the AIP-SFFP plans were similar to the research plans results(24.9±4.3Gy/23.1±2.0Gy/23.9±2.3Gy/22.7±4.8Gy/34.7±2.5Gy). D(0.01cc) results of brainstem and cord+5mm in AIP-SFFP plans were also similar, but body max dose results were slightly higher.


Conclusions:
This novel AIP-SFFP with local-attention design was successfully demonstrated for H&N IMRT application. Future developments make it promising for future clinical applications.

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

Keywords

Treatment Planning, Modeling

Taxonomy

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

Contact Email