MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Semi-Supervised Dose Prediction with Generative Adversarial Learning

D Lam*, B Sun , Washington Univ. in St. Louis, St. Louis, MO

Presentations

(Wednesday, 7/17/2019) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 2

Purpose: Most artificial intelligence applications in radiation therapy are supervised learning where the label collection is very expensive and time consuming. This paper presents a new method for dose prediction that leverages unlabeled data in the treatment planning, which otherwise would be useless since there is no ground-truth label.

Methods: 48 prostate cancer patients treated with IMRT were used to train, validate and test the algorithm. All the patients receive 7920cGy. All the doses, which were calculated by Varian Eclipse, are used as the ground truth. To increase the number of training data by exploiting unlabeled data, another 17 prostate cancer patients treated with proton therapy were used as semi-supervised learning, in which only contouring labels were used to assist the training phase of dose prediction. Generative Adversarial Network approach was used in the prediction algorithm. A Dense-U-Net was used as the generator to predict dose from contouring label input. Another Deep Convolutional Neural Network was used as the discriminator to distinguish whether the input dose is from treatment dose or from the generative network. To assist the discriminator, the mask information was also used beside the dose as the input. For contouring label without dose ground truth, a generated dose was calculated by the generator and then was input into the discriminator to generate an adversarial loss to train the network. After training, 4 unseen patients were used as testing.

Results: The training converges at 150k iterations where the generative loss converges to 0 and the maximum prediction error reduces toward 0. The evolution of the dose prediction for 4 test patients for different axial positions at different iterations shows that the algorithm successfully predicted the 3D dose.

Conclusion: By exploiting the GAN architecture, the proposed approach proves that semi-supervised learning can be used for dose prediction.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email