Mr. Justin Dong is now a top student of the 12th grade at Whitney High School, No. 1 High School in California ranked by US News. He has already received many prizes and awards, including High Honors (Top 50 in US) of 2018 Olympiad Chemistry National Competition, Distinguished Honors in the American Mathematics Olympiad, first place winner of Mathcounts Competition, award winner at National Science Bowl, National AP Scholar, and National Merit Scholarship Finalist. With strong interest in mathematics and biology, he was one of the key programmers in UCSD Medical School AI Team while doing his research internship during summer of 2017, and a coauthor of the team’s research paper which was published in the Feb 22, 2018 issue of Cell. In 2015, he began to enter research by shadowing under a professor at UCLA Medical School, learning about cancer and genetics. Mr. Dong founded the Young Math Enthusiasts Club to promote mathematical problem solving in his community, and received the North American Youth Community Leadership Award for his service. He is fascinated by the beauty of scientific problem solving, and enjoys experimenting in the chemistry lab. In his free time, he plays tennis and chess, together with many other hobbies.
An innovative Image-Based Deep Learning system towards Identification of Medical Diagnoses and Treatment of Diseases
Since my childhood, I have been surprised to discover my talent for math after participating in regional math competitions and winning awards. But more importantly than the results, I saw and enjoyed the beauty in solving problems. More recently, I expanded my passion to other fields of science, such as biology. I have kept this throughout my experiences of doing research last summer.
I have always been interested in the potential for computer algorithms to aid in clinical decisions and medical imaging. As a team member of the UCSD Medical School AI project, and guided by advisors, I have been the lead programmer in developing diagnostic tools based on a deep-learning framework for the screening of patients with common treatable diseases. Applying this approach to a dataset of optical images, our team demonstrates performance comparable to that of human experts in classifying age-related macular degeneration and diabetic macular edema. Our project uses transfer learning with a pre-trained model to reduce the amount of data and time needed for training. Instead of training the neural network completely from scratch, we use a set of network weights pre-trained on the ImageNet dataset, with 14 million images of objects. We retrain the decision-making layer of the neural network to make diagnoses for retinal diseases, as well as normal retinas. The ability to classify ordinary objects is retained in the internal layers of the network, and this knowledge also transfers to the task of classifying retina images. In addition, we confirm that our diagnosis tool correctly identifies the characteristics of each disease using an occlusion test. By running predictions on images with various parts covered, we pinpoint the regions that were most important for the machine’s diagnosis. This allows us to generate maps of probabilities assigned to the correct diagnosis, providing a visualization of how the neural network makes its predictions. Our algorithm can be incorporated in a tool that will allow patients to upload retina images, which will be used to quickly make a computer prediction. This tool might ultimately aid in expanding access to expert diagnoses, thereby facilitating earlier treatment, resulting in improved clinical outcomes.