Screening systems are an essential element of diabetic retinopathy prevention. The efficiency of these systems could benefit from computer-aided approaches. Eye fundus image processing methods are developed in this thesis, in order to detect lesions and produce pathological scores. The work of Xiwei Zhang has been integrated in the TeleOphta system, which achieves a good performance, comparable to human experts.
Diabetic retinopathy is a complication of diabetes. It is the main cause of blindness among the middle-aged population. An early detection and adapted treatment considerably reduce the risk of sight loss. Medical authorities recommend an annual examination to diabetic patients. Doctors examine patients’ eye fundus images to give a grading. The grading is based on the lesions found in the fundus images, as well as on contextual data, as patients’ background. Following this grading a recommendation is issued to the patient, e.g. next screening in 12, 6 or 3 months; to be referred to hospital eye service; etc.
Fundus examination is not performed sufficiently because of massive screening work. The aim of the TeleOphta project is to automatically detect normal examinations in a diabetic screening system, thus to reduce the burden on readers, and therefore serve more patients. The project is funded by the French ANR (Agence Nationale de la Recherche). The partners are: AP-HP (Assistance Publique - Hôpitaux de Paris), which provides medical expertise and databases; LaTIM (Laboratory of Medical Information Processing - University Hospital of Brest, Faculty of Medicine and Telecom Bretagne ) and Centre for Mathematical Morphology, MINES ParisTech - CMM, which are in charge of image processing and data mining methods development; and, last but not least, ADCIS, which is in charge of software integration. The project is coordinated by CMM. The developed system classifies each patient record into two categories: “To be referred” or “Not to be referred”. For more details, refer to http://teleophta.fr/.
The thesis of Xiwei Zhang proposes several methods to extract information linked to diabetic retinopathy lesions from color eye fundus images. The detection of exudates, microaneurysms and hemorrhages is discussed in detail. One of the main challenges of this work is to deal with clinical images, acquired by different types of eye fundus cameras, by different persons. Therefore the data base heterogeneity is high. New pre-processing methods, which perform not only normalization and denoising tasks, but also detect reflections and artifacts in the images, are proposed. Novel candidate segmentation methods based on mathematical morphology, and new textural and contextual features for lesion characterization, are proposed. A random forest algorithm is used to detect lesions among the candidates. The proposed methods make extensive use of new residue analysis methods.
The proposed methods have been integrated within the TeleOphta system, which has been evaluated on two large databases. The classification is based not only on the results of the proposed methods, but also on image signatures provided by other partners, as well as on medical and acquisition-related information. The evaluation shows that the TeleOphta system can make about 2 times more patients benefit from the diagnosis service, based on an existing telemedicine network.
Xiwei Zhang was born in Shanghai, China. In 2008, after obtaining his bachelor degree of telecommunication engineering in Shanghai Tongji University, he came to Telecom Bretagne in Brest, France, for a master degree in image processing. During the master course, he did an internship in the Centre for Mathematical Morphology (MINES ParisTech) in 2010. The subject already dealt with retinal image processing.
Then, he decided to continue studying in this domain by starting a PhD thesis, under the supervision of Etienne Decencière. He obtained his PhD degree in 2014. Now, he is working as a post-doc in the Centre for Computational Biology, a joint laboratory between MINES ParisTech, Institut Curie and INSERM, dedicated to epidemiology, bioinformatics and systems biology of cancer.
|©2009 Mines ParisTech