Room: Exhibit Hall | Forum 9
Purpose: Stereotactic body radiotherapy is an effective treatment modality for non-operable early-stage lung cancer. Although symptomatic radiation pneumonitis (RP) is reported to be less than 10%, with various patient-specific confounding factors, there is still a strong need to augment current dosimetric RP models towards more accurate predictions. This study presents a deep neural network (DNN)-based RP predictive model.
Methods: 66 lung cancer patients underwent SBRT were retrospectively studied. The median duration of follow-up is 10.6 months. Tumor location, fractionation pattern, absolute volumes of the total lung and ipsilateral lung, mean doses (cGy) of lung and ipsilateral lung, V40, V30, V20, and V5 of the total lung were acquired as the input. RP endpoints serves as the output. A binary model and a quaternary model were built based on a five-layer neural network. In the binary model, output 0 and 1 represent â€œno RPâ€? (N=48) and â€œRPâ€? (N=18), whereas in the quaternary model, output 0,1,2,3 respectively represent â€œno RPâ€?, â€œRP detected radiologicallyâ€?, â€œRP detected clinicallyâ€? and â€œRP detected bothâ€?. Both models were built from a training cohort (N=51), which was randomly selected, and validated in the validation cohort (N=15). 10-fold cross-validation were performed 500 times for both models, and the average accuracy were evaluated. Receiver Operating Characteristic (ROC) curve of the model validation and mean ROC curve of the repeated cross-validation were utilized to evaluate the binary model.
Results: The binary model achieved an area under the area (AUC) of 0.91 in the validation cohort (N=15). The average accuracy of cross-validation is 96.7% Â± 6.1%, and 98.5% Â± 2.3% for binary and quaternary model, respectively. Repeated cross-validation achieved a mean AUC of 0.98 Â± 0.06.
Conclusion: Our DNN-based model achieves high accuracy of RP prediction. More patient-specific prediction warrants a larger dataset and other biological/clinical variables.