CircWaveDL: Modeling of optical coherence tomography images based on a new supervised tensor-based dictionary learning for classification of macular abnormalities

Abstract

Modeling Optical Coherence Tomography (OCT) images is crucial for numerous image processing applications and aids ophthalmologists in the early detection of macular abnormalities. Sparse representation-based models, particularly dictionary learning (DL), play a pivotal role in image modeling. Traditional DL methods often transform higher-order tensors into vectors and then aggregate them into a matrix, which overlooks the inherent multi-dimensional structure of the data. To address this limitation, tensor-based DL approaches have been introduced. In this study, we present a novel tensor-based DL algorithm, CircWaveDL, for OCT classification, where both the training data and the dictionary are modeled as higher-order tensors. We named our approach CircWaveDL to reflect the use of CircWave atoms for dictionary initialization, rather than random initialization. CircWave has previously shown effectiveness in OCT classification, making it a fitting basis function for our DL method. The algorithm employs CANDECOMP/PARAFAC (CP) decomposition to factorize each tensor into lower dimensions. We then learn a sub-dictionary for each class using its respective training tensor. For testing, a test tensor is reconstructed with each sub-dictionary, and each test B-scan is assigned to the class that yields the minimal residual error. To evaluate the model's generalizability, we tested it across three distinct databases. Additionally, we introduce a new heatmap generation technique based on averaging the most significant atoms of the learned sub-dictionaries. This approach highlights that selecting an appropriate sub-dictionary for reconstructing test B-scans improves reconstructions, emphasizing the distinctive features of different classes. CircWaveDL demonstrated strong generalizability across external validation datasets, outperforming previous classification methods. It achieved accuracies of 92.5 %, 86.1 %, and 89.3 % on datasets 1, 2, and 3, respectively, showcasing its efficacy in OCT image classification.