Monitor Rentinal Injuries using OTC retinal Imaging

Problem statement

Sickle cell disease (SCD) is an inherited group of hemoglobinopathies. In SCD, organ damage begins with occlusion of small blood vessels by sickled red cells. Interruptions of retinal circulation in SCD result in retinal structure change – sickle cell retinopathy (SCR).

Optical coherence tomography (OCT) is a non-invasive imaging technique that uses light waves to take cross-sectional pictures of the retina with micrometer resolution. OCT reveals each of the retina’s distinctive layers, allowing ophthalmologists to map and measure their thickness. Retinal infarction associated with SCR and other vaso-occlusive diseases causes selective thinning of the inner retinal layers on OCT study.Early detection of the disease could lead to early intervention and improve outcome

Current SCR monitoring protocol: annual retina examination including OCT imaging. Results of OCT study are generated by ophthalmologists visually inspect each image and make comparison to patient’s earlier studies. This is time consuming and may be affected by personal experience, attention and other subjective factors.


Project Goals

Our task is to provide a way to automate detecting retinal thinning found in OCT images to achieve the following:

Establishing "normal" standard

During the process we should have a group of OCT images that are healthy that can be used to compare with other photos

By establishing a "normal" standard we can compare and find what features in an image can be classified as thinning.

50% accuracy or higher

Ultimately the method used should have be able to correctly find thinning in OCT images.

If a specific accuracy cannot be achieved then the method should in some way help the ophthalmologists.

Have high specificty with High specifity

The proposed solution should be able to correctly classify a person with having the disease or not

Sensitivity refers to the model's ability to designate an individual with the disease as positive. Sensitive means that there are few false negative results.

Proposed Solution: Instance Segmenation

Instance segmentation is the task of identify object outlines at the pixel level. The reason for this choice of method is due to the problem at hand. Our main task is to identify sections of thinning that can possibly appear in various OCT images. Since we already know what the image is about theres no need to see this as a classification problem, narrowing our options to object detection or instance segmentation. The secondary tasks are being able to count the number of detected thinning then calculate distance of each from an important feature called the fova center. In order to calculate distance we need to be able to know exactly where the objects of interest are in the image. Instance segmentation overcomes this by detecting objects at pixel level.

For implementation our model is going to be based on Mask RCNN. Mask RCNN is an instance segmentation model that's based on Feature Pyramid Network and a ResNet101 backbone. This is a popular model that has been modified in various other projects to meet their needs and acheive high results. For our case we should be able to do the same.