From Personalized Retinal Image Mapping to Large Scale Parallel Image Processing

Опубликовано 17 августа 2016, 3:05

Being one of the most elegant sensory organs of the human body, anomalies in the retina can have serious consequences on the vision system and are indicative of specific medical conditions. Off the shelf fundus cameras are capable of acquiring high quality retinal images, yet accurate interpretation of retinal images remains a major bottleneck in the early detection of asymptomatic retinal conditions. Decades of research has led to numerous retinal image segmentation algorithms which can map blood vessels in retina images, and identify some of the most telling retina lesions. Despite this significant progress, how to derive personalized optimal parameters for automated mapping remains an open issue. In this talk, I will discuss our recent work in personalized modeling of blood vessel mapping algorithms. The technique is based on statistical fitting of selective image features to a statistical distribution, using a Sobel (or similar) edge detector as a ΓÇ£probeΓÇ¥ to measure the personalized optimal threshold parameter for BV mapping. The approach allows systematic tradeoff analysis between detection sensitivity and false detection rates. The implementation of these mapping algorithms on a Pico FPGA card on a laptop achieves 30x acceleration over a CPU based implementation. We plan to utilize the very high level of parallelism offered by FPGA cluster/FPGA cards to extract region-based features for cloud-level image indexing/retrieval applications, and client crowd-based image feature processing and aggregation.