Abstract
Purpose :
To automatically classify, using a deep learning model, spectral-domain optical coherence tomography (SD-OCT) and blue fundus autofluorescence (FAF) images of Best vitelliform macular dystrophy (BVMD) and adult vitelliform macular degeneration (AVMD).
Methods :
Fifty-two BVMD eyes (SD-OCT images and FAF images) and 43 AVMD eyes (SD-OCT images and FAF images) were included. The contrast-based auto local threshold was used to preprocess FAF images. SD-OCT B-scans were binarized and processed using a mean-based auto local thresholding. MatLab deep learning toolbox was used as a framework for the deep learning process. Images were classified using Inception-ResNet-v2 convolutional neural network(CNN). Transfer learning using the ImageNet dataset was performed. Established augmentation techniques were used. Seventy (70)% of the images were used to train the network while 10% was used for validation and 20% for testing. Accuracy, sensitivity, and specificity were assessed using confusion matrices. The area under the receiver operating characteristics (AUROC) curves was determined to evaluate the model performances. The deep learning model's confidence was assessed using binomial logistic regression on the test set.
Results :
The accuracy of the classification using FAF images was 96.8%. The model showed a 98.0% sensitivity and 95.4% specificity for the diagnosis of BVMD with an AUROC of 0.980 (CI 0.949 – 1.000). The accuracy of the classification using SD-OCT images was 0.936. The model showed a sensitivity of 94.2%, a specificity of 93.0%, and an AUROC of 0.942 (CI 0.890 – 0.994) for the diagnosis of BVMD. Binomial logistic regression revealed a 92% of correct classification probability.
Conclusions :
The study shows good performances of CNN-based automated classification in differentiating the vitelliform stage of BVMD lesions from AVMD based on preprocessed FAF and SD-OCT images. With further developments, this model may help evaluate and distinguish the two clinical entities.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.