Project Summary/Abstract Cancer is the leading cause of death from disease for children and young adults in developed countries. Although primary lung cancer is rare in this population, the lung is a common site for many cancers to metastasize forming pulmonary nodules. Such lung nodules may be identified in thoracic computed tomography (CT) scans. Early and accurate detection of lung nodules in pediatric CT is critical for therapy planning and optimization, correct cancer staging, and disease monitoring. However, identifying pulmonary nodules in CT is an arduous and time- consuming task for radiologists, fraught with considerable inter-reader disagreement. Because of the difficulty of the task, even experienced radiologists may fail to identify potentially significant pulmonary nodules. These challenges are exacerbated in pediatric cases where smaller and more subtle metastatic nodules are likely to be clinically significant. The goal of this project is to develop a pediatric-specific computer aided detection (CAD) system optimized for the detection of clinically meaningful nodules in children. The system will significantly improve the ability of radiologists to detect lung nodules, and ultimately improve outcomes for children with cancer. Despite the need, to the best of our knowledge there are currently no pediatric-specific pulmonary nodule CAD algorithms. Our central hypothesis is that the pediatric-specific CAD system will outperform those designed for adults when applied to pediatric patients. We will evaluate our hypothesis via the following two specific aims. Aim 1 will focus on the development of a pediatric-specific pulmonary nodule CT CAD system using state-of-the-art deep learning models and pediatric training data. Under Aim 2, we will evaluate the diagnostic performance of our pediatric-specific CAD system compared to that of existing adult-trained systems in a pediatric sample, benchmarking against multi-radiologist review. This project will produce what we believe will be the first pediatric-specific CAD system for pulmonary nodule detection, with the ultimate long-term goal of clinical translation to improve outcomes for children with cancer. The proposed effort is in direct support of the Image IntelliGently initiative of the American College of Radiology, which endorses the principle that “AI used in pediatric patients should be designed for and shown to work in pediatric patients.”

Project Narrative Cancer is the leading cause of death from disease for children in developed nations. The overall aim of this project is to develop and test artificial intelligence methods, specifically designed for children, to detect when cancer has spread to the lungs (one of the most common sites of metastatic cancer). Our pediatric-specific system will provide earlier and more accurate detection of lung nodules in computed tomography imagery to help doctors adjust treatment to improve outcomes for children with cancer.