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

In recent years, deep learning methods have excelled in Optical Coherence Tomography (OCT) image classification but demand high computational resources and extensive training data. We propose two effective methods for OCT image classification, combining the strength of deep learning with sparse representation of significant image features for improved detection of retinal abnormalities. The first method, Simplified Deep Dictionary Learning (S-DDL), is based on deep dictionary learning, where the loss function enforces a sparse feature representation of input image patches before classification. The second method uses the Wavelet Scattering Transform (WST), which employs predefined filters in network layers. We compare its performance with the S-DDL method. WST is a convolutional network with predefined wavelet filters that do not need to be learned, requiring lower processing time and complexity. Both methods can be directly applied to raw image data without time-consuming pre-processing, as the first layers act like denoising filters to achieve sparse significant image structures. We assess the methods on the Optical Coherence Tomography Image Database (OCTID), consisting of 572 spectral domain OCT volumetric scans in five categories. Our proposed S-DDL and WST-based methods achieved 97.2% accuracy in diagnosing AMD, MH, and Normal categories of OCT images. Additionally, WST achieved 100% accuracy in diagnosing DR or AMD from Normal images. These results are comparable to state-of-the-art methods. Specifically, our proposed method outperforms other methods in detecting one abnormality.