A Deep Learning Model For AMD Classification Based On Volumetric Macular SD-OCT Scans

Purpose: Using three-dimensional convolutional neural network(3D-CNN)algorithm to construct a deep learning model based on volumetric spectral-domain optical coherence tomography(SD-OCT)scans,for age-related macular degeneration(AMD)classification.Methods: Volumetric macular OCT scans and the corresponding colored fundus photographs were collected from patients who visited Joint Shantou International Eye Center between January 2016 and December 2020.According to AMD clinical classification and Age-Related Eye Disease Study,the macular OCT volumes were classified into two classes: the low-risk “Normal,normal aging changes and early AMD” and the high-risk “Intermediate and late AMD”,and the 3D-OCT dataset was established.A deep learning model based on 3D-CNN was constructed for the binary AMD classification and its performance was evaluated.Results: A total of 1027 patients and 1477 eyes were included in this study.The mean age of the patients was 66.5±10.0 years old.There were 615 males(59.9%)and412 females(40.1%),and the ratio of low-risk eyes/high-risk eyes was 2.3:1.1029 eyes(69.7%)were assigned to the training set,231 eyes(15.6%)to the validation set,and 217 eyes(14.7%)to the test set.The 3D deep learning model showed great performance on the test set,with the accuracy of 0.935,sensitivity of 0.973,specificity of 0.916,F1 score of 0.911,and AUC of 0.9888(95%CI: 0.9770-1.0000).Conclusion: This study constructed a deep learning model using 3D-CNN algorithm,for classifying AMD on volumetric OCT scans,and the model demonstrated excellent grading performance.