Personal tools



Jump to: navigation, search

Title Image Analysis and Machine Learning Platform for Innovation in Diabetic Retinopathy Screening
Reference CMUP-ERI/TIC/0028/2014
PI Carlos Costa
Participants Augusto Silva, Sérgio Matos, Jorge Miguel Silva, Micael Pedrosa
Funded by Carnegie-Mellon - Portugal / FCT
Global funding (€) 552044
RU funding (€) 156504
Starts 2016/04/01
Ends 2019/03/31

Diabetic Retinopathy (DR) is a leading cause of blindness in the industrialized world that can be avoided with early treatment, demanding an earlier diagnosis in a stage where the treatment is still possible and effective. The Portuguese North Health Administration (ARSN) is implementing a mass screening for DR, with the goal of making eye exam of about 75% of identified diabetics, from an estimated diabetic population of 250.000, in the north of Portugal. In this DR screening, the decision relies upon expert ophthalmologists that, by analysing the eye fundus images, detect the presence of DR and grade the pathology severity. This task is enormous as it demands the analysis of every retinal image, among a huge image dataset. It is estimated that 75% of patients don't have any DR manifestation, which creates a burden on healthy subject examinations. Furthermore, there are images with very low quality that preclude their analysis. For DR grading, most of the systems focus on the detection of micro-aneurysms (uAs) for differentiating DR from non-DR based on image analysis or using features extracted from the fundus images followed by classifiers[3]. More advanced Image Analysis (IA) and Machine Learning (ML) techniques are required for the identification of lesions associated with moderate/severe DR stages, such as haemorrhages and hard exudates [6,7,8]. For proliferative DR detection, an automatic system should be able to detect the formation of new vessels [7]. However, the majority of the research is, unfortunately, highly optimized for small image sets compromising the generalization capability for mass screening applications [9]. The validation of DR screening is a key issue that have been addressed by several authors [8,9], with particular attention to the use of appropriate validation sets. SCREEN-DR faces several challenges. The 1st is to automatically evaluate image quality, and consequent removal from the workflow the low quality images. The 2nd is to automatically detect the non-pathological cases. If these 2 challenges are overcome, the ophthalmologists need to analyze only about 25% of cases, an important gain in terms of time-to-decision and efficacy. The 3rd challenge is to automatically grade DR in several scales of the disease severity. INESC TEC and Dr. G. Rohde lab from CMU (Carnegie Mellon University) propose several directions for image analysis innovation in a DR context. Moreover, SCREEN-DR extends the index and retrieval PACS concept at UA (University of Aveiro) to the retinopathy use case, under a new concept of medical imaging collaborative platform. The vision of the consortium SCREEN-DR is to create a distributed and automatic screening platform for DR, based on the state-of-the-art Information and Communication Technologies (ICT), including advanced Picture Archiving and Communication Systems (PACS) management, ML and IA, enabling immediate response from health carers. This multidisciplinary consortium is coordinated by INESC TEC (at U.Porto) and by CMU. The other partners are IEETA (UA), BMD-software (a UA spin-off company), and ARSN, that together with hospitals in the region and the University of Pittsburgh Medical Center (UPMC), offer clinical expertise. First Solutions, a medical informatics company, act as consultant of ARSN. The cross-disciplinarity of ICT research is guaranteed through the different roles of the partners. INESC TEC will be focused on the IA and pattern recognition (PR)tasks for assessment of image quality, for detecting the cases without any signs of pathology and for DR grading. UA will lead the development of a collaborative platform for friendly access to the PACS used by the different players. The CMU team will develop methodologies for evaluating image quality and detecting normal images, by approaching the problem from a ML point-of-view, while the INESC TEC team will embed domain knowledge in the image analysis task. With this complementary approach, we hope to be able to select the best approach or combine if there is any complementarity to be explored. In parallel with the technological tasks, INESC TEC creates an innovation package to ready the innovation to the market by creating a roadmap of a commercialization strategy and prepare to license the resulting technologies. For this purpose, we plan to generate a business model by using Biodesign [25] and customer development [27] processes to define proof-of-concept versions of the imaging modules. This joint collaborative venture of research institutes, together with technology and service providers in the health sector creates an excellent attraction pole for new talented students involved in the multiple educational programs at U.Porto and UA. The research and innovation activities will be interleaved by education actions at PhD (including the dual degree), MSc, and MBA levels. These students will be the core of the project and part of the pillars of the future in ICT in Biomedical Engineering.