QOL-55. INTEGRATED MULTI-SCALE MODEL FOR PEDIATRIC BRAIN TUMOR SURVIVAL PREDICTION

Abstract

Brain tumors are the most common solid tumors affecting children, and its prognosis has been a great challenge for physicians and researchers. With the advances in high-throughput sequencing technology and digital pathology, more quantitative data is now becoming available and more information may potentially be discovered in whole slide images (WSIs) and molecular tumor characteristics to determine survival and treatment. Imaging and genomic data, though very different in nature, both may contain different aspects of disease characteristics that are important for survival prediction. Hence our work aims to build a framework to integrate two data modules, whole-slide histopathology image data, and RNA sequencing data, for a unified model to improve pediatric brain tumor survival outcome prediction. The imaging data and genomic data are both of high dimensions and on different scales. We use two independent modules, each of which consists of a deep neural network, to extract lower dimensional features from imaging and genomic data respectively. We concatenate the extracted features and use a third neural network to train a Cox regression model using the merged feature as input. Each module is first pre-trained with TCGA adult brain tumor data, and subsequently fine-tuned with pediatric brain tumor data. The entire pipeline is tested on the holdout pediatric brain tumor dataset. Preliminary results suggest that the integrated framework achieves improved prediction performance than using each single data module alone. The concordance index (C-index) of integrated model is 0.68, compared to 0.62 with imaging data only, and 0.66 with genomic data only.