Multiple-instance and multi-resolution image mining for diabetic retinopathy screening

Objectives. - Automated analysis of eye fundus images has recently proven useful for diabetic retinopathy (DR) screening. A novel image analysis framework is introduced in this paper for solving two remaining problems shared by all automated DR screening systems: image quality assessment and complex or rare lesion detection. Patients and methods. - A new wavelet-based image description is introduced: given a problem-specific definition of abnormality, this image description is used to search for abnormal areas in images at different resolutions. The proposed search algorithm is based on multiple-instance learning: experts indicate examination records presenting abnormalities, without locating these abnormalities in images, and the algorithm learns to detect them automatically. This algorithm has assessed on a dataset of 25,702 screening examination records, consisting of 107,799 images altogether. Results. - The algorithm was able to detect examination records containing low quality images with an area Az = 0.907 under the ROC curve. It was also able to detect patients with proliferative DR with an area Az = 0.792, as well as differentiating them from non proliferative DR patients with an area Az = 0.819. Discussion - Used jointly with the automated detection of the early signs of DR (small, frequent lesions), the proposed approach should increase the robustness of automated detection of DR and of the other pathologies seen in screening centers. (C) 2011 Elsevier Masson SAS. All rights reserved.