Retinal diseases are the principal causes of low vision and can lead to total blindness. The early detection and diagnosis of these diseases can help to heal them, or slow or limit the damage. This implies that systematic detection means of these diseases are developed.
Due to the decreasing number of ophthalmologists, solutions via telemedicine networks, such as OPHDIAT network led by AP-HP, a project partner, are being implemented.
In order to prevent the ophtalmologists from being snowed under unrelevant files (i.e. without any disease), the project partners intend to develop algorithms for an automatic processing of digital color fundus images. These algorithms, integrated in the ophthalmic telemedicine networks, will implement a classification in images that do not require any expert opinion and images that require a notice, which will include pathological images and unanalyzable images for developed algorithms. The information about the detected diseases will also be transmitted.
The main objective of this program, through the response to a public health problem, is to develop a multidisciplinary research in automatic image processing, pooling two scientific concepts: data mining and digital image processing. The expected result is a method combining both approaches, powerful enough to avoid a significant percentage of images from being read by specialists.
Data mining will be generated from the images and text data presented in a large database provided by the partner AP-HP, from the images collected in OPHDIAT network.OPHDIAT base consists of photographs and records of diabetic patients who participate in a program to detect diabetic retinopathy using telemedicine. The aim will be to determine the "signature" of images/folder in order to decide if the files being reviewed should be sent to specialists or not. This work will include, when they are available, the specific image processing results for the detection of diabetic retinopathy lesions, as well as the information about a possible presence of glaucoma. Improving existing methods of image processing, the search for new methods, is essential for success, considering the vast heterogeneity of image acquisition source.
Digital image processing has two objectives: firstly to extract quantitative information about the lesions related to diabetic retinopathy, to determine the disc optic and look for characteristics indicative of a risk of glaucoma, secondly to study the global descriptors of the qualified images which will help or simplify the images classification, in addition to the determination of local parameters for diabetic retinopathy and glaucoma. This set of descriptors will constitute the digital signature of images. This signature will be integrated in the data mining algorithms that will develop the final decision procedures.
Technically, the developed algorithms will be implemented on OPHDIAT server, in order to carry out an minimum preliminary evaluation, and to prepare a clinical assessment protocol for further evaluation, which is not in the framework of this project.