Room: Stars at Night Ballroom 2-3
Purpose: To build a model using deep learning methods for 18FDG-PET image outcome prediction after oropharyngeal cancer IMRT using pre-radiotherapy (pre-RT) images and planned dose distribution
Methods: 66 oropharyngeal cancer patients undergoing 70Gy IMRT were used to develop the model. 61 patients were used for training/validation and the remaining 5 were used for testing. Each patient received pre-RT PET-CT scan and a 2nd PET-CT scan (intra-RT) after 24Gy delivered in 2Gy/fx. The prediction of intra-RT PET scan was implemented by a modified convolutional neural network (CNN) with 8 convolution layers as a deep learning approach. During network training, axial slices of pre-RT PET and CT images, 24Gy spatial dose distribution and GTV/CTV binary masks were stacked as the inputs, and the registered axial intra-RT PET slices were the outputs. The loss function was a modified mean square error to heavily penalize results in CTV. Raw outputs from CNN were processed by a linear rectification function to harmonize SUV intensity distributions as a fine-tune process. To evaluate prediction accuracy, mean SUV values in GTV/CTV and gamma passing rate of the predicted intra-PET images in reference to ground-truths were reported in 5 test patients.
Results: Intra-RT PET image outcome predictions were successfully generated for 5 test patients. 3D mean SUV values in CTV/GTV by prediction were 1.36/3.01, which were close to ground truth values (1.47/2.91). In gamma analysis (SUV threshold 10% local value, 3mm distance-to-agreement), the average 2D gamma passing rates were 85.8% for all 266 axial CTV-containing slices and 98.8% for all 107 axial GTV-containing slices. Average 3D gamma passing rates in CTV/GTV were 98.8%/99.9%, respectively.
Conclusion: For oropharyngeal cancer patients undergoing IMRT, the proposed model successfully predicted intra-RT PET image outcome with good quantitative accuracy. Current results demonstrate the potential use of the proposed model for dose-paining in RT planning.