Abstract
Purpose :
The purpose of this research is to demonstrate a data-driven feature extraction technique that highlights prominent structures in retinal images for uses in visualization, segmentation, classification, and unsupervised clustering. Data labeling and model building are both time-consuming and expensive processes, and this research aims to expedite the process of initial data triage.
Methods :
We introduce a novel data representation that consists of cropping, restructuring, resizing, and subtracting an image from its original form, as is depicted in Figure 1. Using this image representation, histograms are calculated, reduced via Principal Component Analysis (PCA), and then clustered via t-distributed Stochastic Neighbor Embedding, grouping images of similar characteristics, as can be seen in Figure 2. The entire process relies only on the given dataset in question and does not require any labeling or model building.
Results :
We validate on a set of 1,145 images from a set of diabetic healthcare clinics, graded for quality and pathology by a certified ophthalmic medical technologist. Our data representation enables individual images to be segmented without any filtering or template matching. Our visualization results produce distinct clusters (ranging from “poor” to “high” levels of quality, including other features like bright artifacts, eyelashes, and depigmentation) for real-world sets.
Conclusions :
Retinal images vary due to camera, photographer skill, and individual being photographed. Detecting pathology or quality is often accomplished using large datasets and building models for future inferences. As the factors that influence retinal images change, these algorithms fail to scale. We present a novel approach that simply uses the data itself to perform a self-filtering and clustering with other data in the set. This method works with any camera type and does not require model building or data labeling, a time-consuming and expensive process.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.