Dictionary Learning Informed Deep Neural Network with Application to OCT Images

Medical images are often of very high resolutions, far greater than can be directly processed in deep learning networks. These images are usually downsampled to much lower resolutions, likely losing useful clinical information in the process. Although methods have been developed to make the image appear much the same to human observers, a lot of information that is valuable to deep learning algorithms is lost. Here, we present a novel dictionary learning method of reducing the image size, utilizing DAISY descriptors and Improved Fisher kernels to derive features to represent the image in a much smaller size, similar to traditional downsampling methods. Our proposed method works as a type of intelligent downsampling, reducing the size while keeping vital information in images. We demonstrate the proposed method in a classification problem on a publicly available dataset consisting of 108,309 training and 1,000 validation grayscale optical coherence tomography images. We used an Inception V3 network to classify the resulting representations and to compare with previously obtained results. The proposed method achieved a testing accuracy and area under the receiver operating curve of 97.2% and 0.984, respectively. Results show that the proposed method does provide an accurate representation of the image and can be used as a viable alternative to conventional downsampling.