Dose Predictions for Head and Neck Cancers Using Hybrid Structure Sets Containing Manual and Automated Contours

Abstract

<h3>Purpose/Objective(s)</h3> The purpose of this study was to create a deep learning dose prediction model for head and neck (H&N) cancer treatments using computed tomography (CT) images, planning target volumes (PTV), and hybrid organ at risk (OAR) contour sets containing manually defined structures from historical treatments and artificial intelligence (AI) generated structures. The aim of this work was to enhance dose prediction models by including AI generated contours. <h3>Materials/Methods</h3> 140 H&N patients were selected for model training and testing. All patients had a PTV treated to 6996 cGy and possibly contained one or more of six lower-dose PTVs ranging from 6600 cGy to 5400 cGy. 9 OARs were selected as important critical structures for dose planning and included the total parotids, total submandibular glands, brainstem, cerebellum, mandible, spinal cord, esophagus, oral cavity, and body. In instances where one or more of the OARs was not contoured for the treated plan, an automated contour was inserted in the vacancy. A U-Net architecture was developed for training the dose prediction model and contained 17 input channels including CT images, 7 PTVs, and 9 OARs. Dose was normalized using a maximum dose point such that the objective dose for training the model was between zero and one. The model used five convolution layers with 32, 64, 128, 256, and 512 filters, respectively. After each convolution, 2 × 2 max pooling was performed to down sample the input matrix. 117 patient data sets were used for training and 23 patients were used for testing. Due to computational memory constraints, the model was trained using individual shuffled slices and in batches of 32 slices. Mean square error (MSE) was used as the loss function and a linear activation function was used for generating the predicted dose. <h3>Results</h3> The predicted mean dose to PTV₆₉₉₆ was 1.9 ± 2.7% lower than the actual mean dose. The mean percent differences between the predicted and actual D_95% and D_2% values for PTV6996 were -8.1 ± 5.1% and 1.5 ± 2.9 %, respectively. The maximum predicted dose to the spinal cord was 11.0 ± 10.8% lower than the actual dose. The percent difference and deviation for the mean dose to the total parotids was 15.0 ±17.7% lower than the actual dose. The mean predicted dose to the external body contour was 5.8 ± 10.1% lower than the actual dose. <h3>Conclusion</h3> The model was able to predict a dose distribution that agreed with treatment planning principles for H&N cancers, with target volumes receiving high dose and OARs receiving lower doses. The current model is a work in progress. More work is needed to validate and refine the prediction accuracy, which predicted lower doses to structures than the delivered dose. After refinement, the model will be used to assess the effects of AI generated contours on the predicted dose.