Interpretation of optical coherence tomography by artificial intelligence


Modern ophthalmology faces a growing trend of ocular pathologies associated with bad habits, poor diet and increasing life expectancy of the population. The pressure on medical institutions requires finding an optimal technical solution that would speed up the processing of optical coherence tomography (OCT) images.

Artificial intelligence-based platforms such as Retina AI can help ophthalmologists – their case is exemplary in processing data from patients with the diseases of age-related macular degeneration, diabetic retinopathy, central serous chorioretinopathy, and the presence of epiretinal membrane.

The algorithm being tested was required to segment obvious signs of cystic macular edema, choroidal neovascularization, central serous chorioretinopathy, and epiretinal membrane on retinal tomographic images. The features were defined as intraretinal cysts, subretinal fluid, subretinal hyperreflective material, retinal pigment epithelium detachment, epiretinal membrane, and retinal drusen.

A total of 3500 scans of tomography images were taken for AI training. The size of the validation database was 650 images. All pathologies from the scans were pre-identified by ophthalmologists for future reconciliation.

The basis for the neural network architecture was EfficientNetB0 + FPN. Based on the masks extracted during feature segmentation, the coordinates of the points of the contours of the found features are determined and the parameters of these contours are estimated. The latter can be used to provide recommendations for differential diagnosis in the patient.

How clearly the model can work was determined during the development by the compliance of data processing with the given criteria: true/false positives, true/false negatives – correct or incorrect positives or negatives, the measures of accuracy were chosen to be precision of response and specificity.

In the validation process, a set of CT scans was run through a trained neural network, counting true positives, false positives, true negatives, false negatives, and calculating Accuracy, Precision, Specificity, and AUC (area under the ROC curve, indicating the quality of the neural network’s predictions).

As a result, different accuracy interpretations were given for different signs of pathologies: 98.06% for intraretinal cysts, 96.27% for subretinal fluid, 92.84% for retinal pigment epithelium detachment, 95.52% for subretinal hyper-reflective material, 88.97% for epiretinal membrane, and 89.12% for retinal drusen.