Classification of Malarial Retinopathy Images Using Deep Learning
The detection of malarial retinopathy is a candidate diagnostic test for cerebral malaria. Malarial retinopathy consists of a set of retinal abnormalities like retinal whitening that is unique to severe malaria and common in children with cerebral malaria. Its presence and severity are related to risk of death and length of coma in survivors. Severe malaria is commonly misdiagnosed in Africa, leading to a failure to treat other life-threatening illnesses. Using deep learning, our framework has the potential to become a powerful new tool for studying malarial retinopathy, and other conditions involving retinal leakage.
The detection of leakage in retinal fluorescein angiogram(a type of imaging commonly used in ophthalmology clinics that provides a map of retinal vascular structure and function by highlighting blockage of, and leakage from retinal vessels) images is important for the management of a wide range of retinal diseases. Our framework can automatically detect three types of leakage(large focal, punctate focal, and vessel segment leakage).The effectiveness of this framework was tested by applying it to images from 20 patients with large focal leak, 10 patients with punctate focal leak, and 10 patients with vessel leakage.
Our model, malariaMED, is a convolutional neural network that inputs a digital fundus image and outputs the probability of of a leakage along with a heatmap localizing the areas of the image most indicative of malarial retinopathy.
A large, prospective autopsy study of children dying with cerebral malaria found that malarial retinopathy was better than any other clinical or laboratory feature in distinguishing malarial from non-malarial coma. However, visualization has to date relied on specialist examination techniques.With automation at the level of experts, we hope that this technology can improve healthcare delivery and increase access to medical imaging expertise in parts of the world where access to skilled ophthalmologist is limited.