Exploring Clinically-relevant Image Retrieval for Diabetic Retinopathy Diagnosis

Project Final Report (PDF, 1.69 MB) Disclaimer

Project Details -
Completed

Diabetic retinopathy (DR) is a common cause of blindness. Despite advances in diabetes care, visual impairment is still a potentially devastating complication. Studies show that timely DR diagnosis and treatment can significantly reduce the risk of severe vision loss. Digital retinal imaging has quickly become an alternative to traditional face-to-face evaluation; however it is labor intensive and prone to error or reviewer fatigue. For this reason, researchers explored automated detection and evaluation of diabetic retinal lesions. Potential benefits of automated DR diagnosis include improved consistency and speed over human reviewers. However, clinicians remain superior in detecting and assessing the severity of DR over computer-based systems, which fail to incorporate the experience and variables that clinicians are able to apply in their assessments. More effort is therefore required to improve the performance of such systems.

The specific aims of this study were to:

  • Develop a content-based retrieval system for referencing diabetic retinal images to improve diagnosis 
  • Develop a prototypical DR image management system to improve reviewers’ diagnostic performance 

This project developed innovative machine learning-based algorithms to retrieve clinically relevant images to facilitate timely and accurate evaluation of DR. Images were considered clinically relevant if they contained the same types of lesions with similar severity levels. While previous work in this field used computational approaches that mimicked clinicians’ diagnostic processes, this approach used feature extraction to detect image features appropriate for distinguishing DR lesions. A retrieval interface was also developed to support the deployment of the retrieval engine in a user-friendly fashion.

The retrieval system was evaluated by examining whether the retrieved images have similar DR severity to images stored in a database. On average, the machine-learning algorithms outperformed other approaches in the literature.