Browse Prior Art Database

Retinal Image Quality Classification Using Convolutional Neural Networks

IP.com Disclosure Number: IPCOM000245601D
Publication Date: 2016-Mar-21
Document File: 4 page(s) / 98K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a system for retinal image quality determination that takes as input a digital fundus of the retina and then outputs a decision as to whether it is gradable. The system uses neuro-biological models such as CNNs that imitate the working of the human visual system (HVS).

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 34% of the total text.

Page 01 of 4

Retinal Image Quality Classification Using Convolutional Neural Networks

Image quality assessment is an important step in retinal image assessment for diagnosis of diseases such as diabetic retinopathy (DR), glaucoma, and age related macular degeneration (AMD). Rapid and accurate assessment and easy accessibility of facilities to clinicians are critical factors in the success of large -scale disease screening systems. Digital fundus photography is a commonly used and effective non-invasive examination medium allowing for automated evaluation of the retinal images. It has the potential to reduce the workload of ophthalmologists and increase the cost- effectiveness of screening systems. Reliable automated diagnosis requires images to be of a minimum quality to facilitate relevant feature extraction . Figure 1 shows examples of ungradable images taken from different datasets .

Retinal image quality is impaired by a number of factors , which degrade it to the point of making it ungradable. Parameters such as focus and clarity, field definition, visibility of the macula, and visibility of optic disc are very important for the correct evaluation of retinal image quality. The parameters identified by the Atherosclerotic Risk in Communities (ARIC) [1] study as being the most important can be divided into two major categories. The first category is generic image quality parameters, such as focus and clarity, absence of artefacts caused by haze, dust and dirt, eyelashes, improper cleaning of the camera lens, and total eye blink. The second category is structural quality parameters such as field definition, visibility of the optic disc, and visibility of the macula.

Image quality determination impacts accurate diagnosis, which is extremely important for the success of any screening program. Medical image quality assessment has not been much explored although it is important for automated systems , as many studies report a significant percentage of acquired study images (4-17%) to be of insufficient quality for an automated assessment [1]. Poor quality images have to be discarded,

which is a waste of time and effort. Moreover, it also affects patient diagnosis leading to multiple visits for image acquisition. Therefore, it is essential to have a fast algorithm to identify poor quality images, allowing a quick second acquisition of the patient image.

This disclosure describes a system that takes as input a digital fundus of the retina and then outputs a decision as to whether it is gradable. It uses trained convolutional neural networks (CNN) for feature extraction which are known to imitate the working of the human visual system.

The novelty of the system is the use of CNNs that imitate the working of the human visual system (HVS), because determining image quality is highly dependent on the

working of the HVS. Existing methods for image quality classification use different feature maps but do not solve the problem from a neurobiological per...