[Brainmap]: James Brown - DeepROP: Automated diagnosis of plus disease in retinopathy of prematurity

Wednesday, November 1, 2017 - 13:00 to 14:00
149 13th Street (Building 149), Room 2204


James Brown, PhD 

Retinopathy of prematurity (ROP) is a leading cause of childhood blindness worldwide, but is treatable if diagnosed early. Severe ROP is characterized by “plus disease”, defined by tortuous and dilated retinal vessels. However, clinical plus disease diagnosis is highly variable among clinicians. We have developed an algorithm based on deep convolutional neural networks (CNNs) for fully-automated plus disease diagnosis from retinal images. Our model was trained and evaluated using a multi-institutional dataset of 6,000 retinal images with disease labels from four experts. We evaluated our algorithm using cross-validation, receiver operating characteristic (ROC) analysis and quadratic-weighted kappa coefficients. We show that the algorithm can diagnose plus diagnose with comparable or better proficiency than ROP experts, as well as previous automated methods. Preliminary results from a small dataset of treatment-requiring ROP patients indicate the method has great potential as a treatment monitoring tool. We now look to evaluate our algorithm prospectively on a larger clinical dataset, and develop a low-cost, portable retinal imaging system for use in developing countries.

About the Speaker:
Dr. Brown's research involves the development and application of shape and appearance models to problems in biomedical imaging. He has collaborated closely with biomedical researchers and clinicians in rheumatology, radiology and ophthalmology to better quantify and understand disease from medical image data. He is particularly interested in the development of algorithms for image segmentation, automated diagnosis, anomaly detection, and establishing genotype-phenotype relationships.