Retinal color fundus photography (CFP) is an imaging technique for capturing detailed images of the back of the eye, including the retina, optic nerve and blood vessels. CFP is non-invasive and offers a safe and painless means to diagnose and monitor a variety of eye conditions such as retinopathy, glaucoma and macular degeneration. High quality retinal images are critical for accurate diagnoses and automated analyses; however, current systems often suffer from quality glitches due to systemic imperfections or operator/patient factors. Fundus image enhancement typically entails a one-to-one mapping between a low-quality image and its high-quality counterpart. Techniques are need to improve low-quality CFPs and aid in disease diagnosis and screening tools.
Researchers at Arizona State University and collaborators have developed a context-informed optimal transport (OT) learning framework for handling unpaired fundus image enhancement. As opposed to conventional generative image enhancement methods, this learning framework better preserves local structures while minimizing unwanted artifacts. This framework was derived using the earth mover’s distance and when tested on a large-scale dataset, it demonstrated superiority compared to several state-of-the-art supervised and unsupervised techniques.
This innovative context-aware OT framework leverages deep feature spaces enabling more accurate fundus image enhancement and better disease identification and screening tools.
Potential Applications
- Retinal color fundus image enhancement
- Disease identification such as retinopathy, glaucoma and MD
- Segment blood vessels
- Help to develop automated tools to screen for neurological disorders such as AD and systemic conditions such as diabetes
Benefits and Advantages
- Evaluation across three large retinal imaging datasets demonstrates that this is superior to SOA supervised and unsupervised methods with regards to signal-to-noise ratio, structural similarity index and two downstream tasks
- Minimizes undue excessive tampering to lesions and structures while effectively removing noise
- Strong foundation for genera image enhancement tasks
- Ensures that the transport costs reflect the intrinsic geometric and contextual properties or the data in the deep feature space
- Has the potential for broader application in medical image enhancement including OCT and endoscopy images
- Eliminates the requirement of paired image datasets
- Maximizes the preservation of thinner blood vessels and lesion formation
For more information about this opportunity, please see
For more information about the inventor(s) and their research, please see