Project Summary and Abstract Radiation therapy (RT) has a proven record of efficacy in treating many forms of pediatric brain tumors. However, it is associated with long-term side effects due to damage to surrounding healthy tissue. This is especially important in the pediatric developing brain, where long-term deficits can be seen in cognitive development. To mitigate these deficits, there has been a shift from whole brain irradiation to more targeted treatment by using dose painting intensity modulated radiation therapy. However, to use these techniques, more information is needed about how giving radiation to normal brain structures, called organs-at-risk (OARs) affects outcomes, both in terms of brain anatomy and function. Voxel Healthcare LLC (VH) is the developer of ClickBrain – an automatic pediatric MR brain segmentation tool that uses cloud-based deep learning (Google TensorFlow) technology for radiology clinical decision support. Since the inception of this software, we extended ClickBrain to ClickBrain RT – a system that will combine ClickBrain’s pretreatment brain structure segmentation outputs with radiation planning computed tomography (CTs) and/or magnetic resonance (MRs) imaging to calculate dose to OARs. ClickBrain RT also segments longitudinal MRIs to track outcomes via volumetric changes. In previous projects, we correlated input parameters such as OAR dosing, demographics, tumor type and grade, OAR and tumor volumetric measurements to OAR volumetric outcomes, using a database of brain MRIs from germ cell tumor patients. Here in Aim 1a, we will modify our prediction using a hybrid deep learning and machine learning method for the prediction of treatment outcomes for OAR. We will also be improving our autosegment algorithms. The tumor images extracted from MRI and CT images will be concatenated with the patient specific demographic information (age and gender) for the training of a hybrid neural network architecture. Imaging features will be sent to an autoencoder convolutional neural network (CNN), while the demographic information will be sent to a densely connected multi-layer perceptron (MLP) network. After individual feature extraction and reshaping, they will be concatenated and sent to a MLP model for treatment late cognitive effects prediction. Missing data points will be synthesized using a modified Pix2Pix Generative Adversarial Network (GAN) method. 3. In Aim 1b, we will build on our current interface prototype, which uses clinical and demographic variables (age, chemotherapy dose, tumor type, grade and location) and baseline imaging as input and outputs predicted outcomes for OARs. We will upgrade the interface to provide full functionality and compatibility with existing commercial software for ease of porting files between systems. Our extended validation in Aim 2 will focus on an existing large database from a broad range of brain tumor patients acquired as part of standard-of-care and previous studies at Children...