Room: AAPM ePoster Library
To investigate the feasibility and performance improvement of introducing survival convolutional neural network (SCNN) into locoregional recurrence (LR) prediction from radiotherapy dose distribution, CT and PET for head and neck (H&N) cancer cases.
The VGG-19 CNN and the Cox proportional hazards regression model were combined to establish the SCNN framework. Four SCNN models were trained by inputting the slices with the maximum GTV pixels of dose distribution, CT, PET and the integration of these three matrices, respectively. The model output was the patient LR risk. A cohort of 237 patients with H&N cancer was obtained from The Cancer Imaging Archive. 141 were used to train the models and the other 96 were to validate. Inputs were reinforced with flipping and rotating, and 1000 epochs were used with each minibatch containing 30 patient data during model training. A negative partial log-likelihood was used for the loss function and an adam was chosen to solve it. Model performance was measured by C-index and the Kaplan–Meier curves analysis, and compared with the traditional radiomics-based model.
In the validation set, the C-index of SCNN-based models were significantly higher than that of radiomics-based models for model_CT (0.61 vs. 0.54, p<0.05) and model_CT+PET+dose (0.70 vs. 0.66, p<0.05); and it was equivalent for model_PET and model_dose (0.60 vs. 0.59, 0.60 vs. 0.60, both p>0.05). Furthermore, for SCNN-based models, model_PET and model_CT+PET+dose, could successfully differentiate the Kaplan–Meier curves of high- and low-risk groups with p<0.05; but for radiomics-based models, only model_CT+PET+dose could realize this differentiation.
The SCNN models were established with the capability of automatically extracting features from dose distribution, CT and (or) PET images, and it can improve predicting LR for H&N cancer compared to traditional radiomics models.
Funding Support, Disclosures, and Conflict of Interest: The present work was supported by the Natural Science Foundation of Guangdong Province, China (No. 2018A0303100020), National Key R&D Program of China (Grant NO. 2017YFC0113203) and the National Natural Science Foundation of China (Grant NO. 11805292, 81601577and 81571771).