Author: Ibtisam Almajnooni, Elisabeth Weiss, Lulin Yuan 👨🔬
Affiliation: Virginia Commonwealth University 🌍
Purpose: We developed a deep learning neural network (DLNN) to predict the risk of radiation-induced esophagitis (RE) during lung cancer radiation therapy based on the spatial dose distribution, for the purpose of identifying critical esophageal sub-volumes and associated dose limits.
Methods: A dataset of 90 locally advanced lung cancer patients was used to train a 3D convolutional neural network to predict the probability of high grade RE (grade>=2). The grade of RE was assessed prospectively based on CTCAE v5 criteria. The network architecture was optimized to represent the serial organ functionality as well as the parallel structure component of esophagus. We utilized anisotropic convolutional kernels with different sizes in the cross sectional and longitudinal dimensions. The kernel dimensions vary at different stages of the convolutions with initial convolutions mainly along the cross sectional and then gradually changed to longitudinal dimensions. The 3D dose distribution, CT image and the structure contours (esophagus, heart, lung and PTV) were combined as input to the network. The dataset was split into 60, 10 and 20 for training, validation and testing. To increase the model sensitivity for the prediction of RE, weighting factors were applied to the training data and the factor values were optimized during hyper-parameter tuning.
Results: The best Area Under ROC curve (AUC) of 0.83 was achieved with the relative weighting factor of 2 between the RE and non-RE training cases. Using a probability threshold of 0.5 for high grade RE, the network identified 3 true positive cases correctly from 4 positive cases within the 20 test cases, with sensitivity: 0.75, specificity: 0.9375 and misclassification rate: 0.1.
Conclusion: A DLNN can be used to improve the prediction of the risk of RE utilizing sub-volume dose distribution. Further studies will incorporate patient clinical factors and validate the model on larger dataset.