MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

An Automated Method to Generate Patient-Specific Dose Distributions for Radiotherapy Using Deep Learning

X Chen*, K Men , J Yi , Y Li , J Dai , National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences, Beijing, Beijing

Presentations

(Monday, 7/30/2018) 4:30 PM - 6:00 PM

Room: Karl Dean Ballroom C

Purpose: We developed an endâ‚‹toâ‚‹end Intelligent Dose (Idose) method to predict patient-specific dose distribution for radiotherapy planning based on prior knowledge with deep learning. Suitable convolutional neural networks (CNNs) with reasonable inputs and outputs were proposed and tested for the task.

Methods: Structure images (StrutImgs) with contoured structures corresponding to CT images were generated as the inputs of Idose, which reflected the precise spatial position information of organs at risk (OARs) and targets. To reduce the complexity of origin dose maps, condensed dose maps (CDMs) were generated as output for CNNs. The final outputs of fine dose maps (FDMs) were generated from CDMs by convoluting smooth filters. Two networks adapted from VGGNet (16 layers) and ResNet (101 layers) were used for modeling, respectively.80 patients with early-stage nasopharyngeal cancer (NPC) were used in this study, of which 70 cases for training and 10 cases for testing. The accuracy of predicted dose distribution was evaluated with the voxel-based mean absolute error in the range of outline (MAEoutline) and a global three-dimensional gamma analysis.

Results: Both Idose models could generate patient-specific dose distribution accurately. Idose with ResNet performed relatively better in MAEoutline than VGGNet (5.5±6.8% vs. 6.2±8.2%, p<0.01). 3D dose distribution predicted with ResNet showed a higher Gamma pass rate than VGGNet. Some small organs (bilateral lens, bilateral optic nerve, optic chiasm) had inferior pass rates (77.7–82.3%) with criteria of 5%/4 mm, but the dose distribution of most OARs could be predicted accurately using ResNet with pass rates >85.0%.

Conclusion: Our Idose model with ResNet is a robust method to generate patient-specific dose distributions for radiotherapy. It can be applied on a new planning CT with contoured OARs and targets to obtain the dose distribution slice-by-slice for planning quality assurance and automated planning.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email