DiscoRet – Automatic discovery of pathological features in retina with unsupervised deep neural networks

Name of the participant: Nishant Kumar

Project description: The development of Optical Coherence Tomography (OCT) and fundus photography has enabled the high resolution and non-invasive imaging of the human eye’s retina thereby allowing clinicians to identify retinal diseases such as diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma. These retinal diseases, if undetected for a long time can cause blindness in a significant proportion of the world’s adult population. The progression of retinal diseases into vision loss could be slowed if the disease is timely detected. A clinical expert could examine the retinal images for the presence of lesions during the initial stages of the disease to prevent its progression to severe levels or even vision loss. However, the detection of the retinal lesions with the help of manual annotations is a time-consuming process with high demands of clinical expertise and equipment. Additionally, the availability of human experts is insufficient especially in underdeveloped countries with high rates of diabetes in the local population due to which low-cost retinal detection is of utmost importance. Therefore, the need for a comprehensive method which has negligible manual retinal screening requirement has long been recognized.
This project deals with the development of an automatic discovery method to detect retinal lesions using clinically acquired retinal images based on unsupervised learning-based neural networks. To this end, different implementational possibilities are to be explored, analyzed and solutions are to be conceived to gain high performance and accuracy of the developed method. For achieving this, an extensive evaluation will also be accomplished with test cases for detecting even those retinal lesions which are extremely difficult to find by a naked eye. Also, the solution will be presented in a manner that it can be generalized to varied camera specifications and acquisition settings. One of the priorities will be to transfer the knowledge obtained during the project to the industry partner for real-world applicability.​

Software Campus partners: TU Dresden, Carl Zeiss Meditec AG

Implementation period: 01.04.2021 – 31.03.2023