TELEmedicine in OPHTAlmology for pathological images automatic detection and assistance in diabetic retinopathy diagnosis – TELEOPHTA

Retinal pathologies are the main reasons for visual impairments that may lead to total blindness. Early detection and diagnosis of these pathologies can help to cure, slow-down or even limit the impact of these pathologies on the sight. This detection requires to setup a systematic screening of these pathologies.
Due to a reduction of the number of ophthalmologists, it becomes very interesting to build innovative solutions based on telemedicine, built upon a network of image acquisition sites and expert centers. Such networks are starting to emerge in various countries. To prevent ophthalmologists from examining patients with no pathology, the project partners are proposing to develop algorithms to automatically process color fundus images.
These algorithms will be integrated in ophthalmologic telemedicine networks.

They will primarily sort images into two categories, images that do not require a review by the specialists (normal images: IN in the text below) and images that require a review by specialists. The latter will include images with pathologies and images that cannot be processed by the imaging system (Image for review: IPA in the text).
One of the key innovative features of the programme consists in using the full information for each patient, including images and patient data. This information will help to take the decision to ask a specialist to review the images along with the patient data.
The programme will use data and images acquired in the diabetic retinopathy screening network called OPHDIAT of the French AP-HP.

The project development will include the extraction of numerical features in the images, first the detection of local features due to diabetic retinopathy, the target pathology of the OPHDIAT network, and the automatic extraction of the optic disk (glaucoma detection), and also the computation of global information on the whole image (signature).
It does not seem reasonable to try to detect all possible pathologies. It also seems difficult to detect acquisition defects that cause images not to be processable by the software, even if they are not pathological images. The determination of these global image signatures should provide some information on the normal or abnormal visual aspect of the images.
Numerical data will be used in conjunction with the patient data to help the decision process. Past experience shows that the use of such information help the specialist to make a decision.
To define the decision process (classification of patient images/files in IN and IPA categories), methods based on data mining will be used on the OPHDIAT database. Along with this information, results obtained by image processing techniques (RD and glaucoma) will be incorporated into the data given to the specialist who will give a diagnosis.
Team members of the project, ARMINES-CMM, the LaTIM (Telecom Bretagne), the ophthalmology department of Lariboisiere hospital (AP-HP), and ADCIS already worked on a programme for the screening of diabetic retinopathy based on retinal images. They will share their expertise on image processing techniques and data mining in this project.