Zheng Y, Williams BM, Pratt H, Al-Bander B, Wu X, Zhao Y. Computer aided diagnosis of age-related macular degeneration in 3D OCT images by deep learning. Investigative Ophthalmology & Visual Science. 2017 Jun 23;58(8):824-.
https://iovs.arvojournals.org/article.aspx?articleid=2638579
Abstract
Purpose
Three-dimensional (3D) optical coherence tomography (OCT) images are increasingly used in the management of eye disease, yet there has been no corresponding increase in the availability of software tools to support the analysis of large amounts of 3D OCT data. We propose a new computer aided diagnosis (CADx) model based on deep learning for automatic diagnosis of eye disease in 3D OCT images and demonstrate its performance with an application to the diagnosis of age-related macular degeneration (AMD).
Methods
384 3D spectral domain (SD)-OCT images (269 intermediate AMD and 115 normal eyes, one image per eye) from a public dataset provided by Duke University were used. 324 randomly chosen images (228 AMD and 96 normal) were used to train the deep learning CADx model while the remainder were reserved for testing the model. The VGG-M network was adopted here for the purpose of classification which comprises 5 convolution layers, 5 max pooling layers, and 3 full connection layers. A cross entropy function was used as the cost function to train the network. Drop-out (ratio 0.5) was used in order to reduce the problem of overfitting during training. The training images were augmented by applying horizontal flipping and random rotation (range 0-10 degrees) to the original image in order to improve classification performance. A class weight (1 to 3 with a step of 0.5) was applied to the normal class to alleviate classification problems associated with imbalanced datasets. Sensitivity, specificity and accuracy of classification were used to evaluate the performance of the trained model.
Results
The sensitivity, specificity and accuracy of the best model with data augmentation and class weighting were 0.927, 0.821 and 0.893 respectively. These results were significantly higher than those of the correspondence model without data augmentation (p<0.0001). The specificity (resp. sensitivity) value increases (resp. decreases) with the increase of the class weight. The model with a weighting value 2.5 yields the best accuracy (p<0.001 when compared to the model without class weighting).
Conclusions
Our results have demonstrated that deep learning based CADx models can provide very encouraging classification performance for the diagnosis of AMD. The developed model could be further developed and validated with large datasets in order to support the management of eye disease.
