MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Region Specific Dose Prediction Using Deep Neural Networks: A Feasibility Study On the Planning Target Volume of Prostate IMRT Patients

D Nguyen*, S Jiang, Medical Artificial Intelligence and Automation (MAIA)Laboratory, UT Southwestern Medical Center, Dallas, TX

Presentations

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

Room: AAPM ePoster Library

Purpose: Deep learning-based volumetric dose prediction models may be hindered by an inherent smoothness and continuity constraint on the dose distribution. Prediction errors from one region may be propagated to another region by such a constraint. We investigate whether a volumetric dose prediction model, using deep learning, on the whole body or only on the planning target volume (PTV) has any effect on the model’s performance on the PTV.


Methods: 72 prostate patients treated with intensity modulated radiotherapy, divided into 57 training, 5 validation, and 10 test, were used. We investigated 2 deep learning models: 1) Body Model: predict the entire body’s 3D dose distribution, and 2) PTV Model: predict only the PTV’s 3D dose distribution. To maintain fairness, both models used the entire volume of 128 x 128 x 32 voxels at 5 mm3 voxel size. In addition, both models utilized identical U-net style architectures, the same optimization algorithm Adam, learning rate of 1×?10?^(-3), and 200 epochs of training. For the PTV Model, the loss outside the PTV was set to 0.


Results: On the 10 test patients we evaluated the PTV mean dose, max dose, dose coverage, and homogeneity. The PTV Model outperformed across every metric compared to the Body Model. Specifically, the PTV model had a prediction error of 0.86±0.29%(Dmean), 0.99±0.93%(Dmax), 1.22±0.40%(D95), 1.25±0.54%(D98), 1.7±0.61%(D99) and 0.0105±0.0078(homogeneity). The Body model had a prediction error of 2.30±1.51%(Dmean), 1.79±1.60%(Dmax), 3.68±1.10%(D95), 4.20±1.20%(D98), 4.76±2.47%(D99) and 0.0258±0.0136(homogeneity). Except for Dmax(p-value=0.0787), all other differences were statistically significant(p-value<0.05).


Conclusion: Focusing a dose prediction model to learn a particular region of interest may significantly improve the performance of the model. By only allowing a loss value for inside the PTV, the model no longer has concerns for PTV boundary’s smoothness and continuity constraints as before, and potential error outside the PTV is no longer propagated.

Download ePoster [PDF]

Funding Support, Disclosures, and Conflict of Interest: National Institutes of Health (NIH) R01CA237269

Keywords

Intensity Modulation, Treatment Planning, Dose

Taxonomy

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

Contact Email