Skychain AI for retinal diseases recognition
Greetings, Skychain community! We have launched the sixth ANN in Skychain Alpha available for public testing — it’s the AI for retinal diseases recognition.
The ANN for retinal diseases detection uses retinal optical coherence tomography (OCT) images to recognize a disease.
OCT is an imaging technique used to capture high-resolution cross sections of the retinas of living patients.
Skychain used the optical coherence tomography (OCT) dataset available on Kaggle. The images it contains were selected from retrospective cohorts of adult patients from the Shiley Eye Institute of the University of California San Diego, the California Retinal Research Foundation, Medical Center Ophthalmology Associates, the Shanghai First People’s Hospital, and Beijing Tongren Eye Center between July 1, 2013 and March 1, 2017.
There are 84,484 retinal OCT images classified in 4 categories:
- NORMAL — images of normal retinas;
- CNV — images of choroidal neovascularization;
Choroidal neovascularization involves the growth of new blood vessels that originate from the choroid through a break in the Bruch membrane into the sub–retinal pigment epithelium or subretinal space. CNV is a major cause of visual loss.
3. DME — images of diabetic macular edema;
Diabetic Macular Edema is an accumulation of fluid in the macula — part of the retina that controls our most detailed vision abilities — due to leaking blood vessels. It is one the most prevalent causes of visual loss in industrialized countries.
4. DRUSEN — images of optic disc drusen.
Optic disc drusen are globules of mucoproteins and mucopolysaccharides that progressively calcify in the optic disc. They may be associated with vision loss of varying degree occasionally resulting in blindness.
Model of the ANN
A modiﬁed version of the VGG16 convolutional neural network was used as the deep learning model for classiﬁcation. We replaced the ﬁnal fully-connected layer with one that has four outputs. We also tried others models as DenseNet169, ResNet, Inception-v3, ShufﬂeNet but no one of them was good as VGG16. The weighted cross entropy was used as the loss function.
After training the deep learning model, we achieved an accuracy of 96,88% on training dataset and an accuracy of 99,9% on validation dataset. After constructing an ROC curve, the area under the ROC curve (AUROC) was 99%.
Below there are some examples of the ANN predictions using retinal OCT images. It creates heat maps to highlight questionable areas.