Application of Deep Dictionary Learning and Predefined Filters for Classification of Retinal Optical Coherence Tomography Images

Computer-aided diagnosis (CAD) has played an important role to help ophthalmologists the diagnosis of retinal abnormalities. Among CAD methods, deep learning methods have high performance in classification. On the other hand, an efficient image representation is required for classification. Dictionary learning is proved to be very effective in finding the best descriptive atoms of an image. Deep Dictionary Learning (DDL) combines the strengths of dictionary learning with deep learning. In this paper, we study two approaches for the detection of retinal abnormalities from Optical Coherence Tomography (OCT) images. In the first method, called S-DDL, we propose a way to avoid the vanishing gradient in DDL and to decrease the training time by simplifying the DDL's calculation. The second method is based on the Wavelet Scattering Transform (WST), which employs predefined filters in network layers. We compare its performance with the S-DDL method. WST is a particular convolutional network in which filters of the layers are predefined wavelets and do not need to be learned. Therefore the WST requires a lower processing time and complexity. For both methods, we do not use any preprocessing on data to obtain an efficient model and decrease computational costs. We assess the methods on Optical Coherence Tomography Image Database (OCTID). This database consists of more than 500 spectral domain OCT volumetric scans, containing five categories. Our results show that the proposed S-DDL method has a high accuracy for OCT image classification. However, the WST achieves better classification results, especially for more than three classes.