Modeling optical coherence tomography (OCT) images is highly beneficial for various image processing applications as well as assisting ophthalmologists in the early detection of macular abnormalities. Sparse representation-based models, particularly dictionary learning (DL), play an important role in image modeling. Traditionally, DL transforms higher-order tensors into vectors and aggregates them into matrices, disregarding the multi-dimensional inherent structure of data. To overcome this problem, tensor-based DL approaches have been developed. In this study, we propose a tensor-based DL algorithm named CircWaveDL for OCT classification where both the training data and the dictionary are higher-order tensors. Instead of random initialization of the dictionary, we suggested initializing it with CircWave atoms, which has previously demonstrated its effectiveness in OCT classification. This algorithm employs CANDECOMP/PARAFAC (CP) decomposition to factorize each tensor into lower dimensions. Subsequently, we learn a sub-dictionary for each class using the training tensor of that class. A test tensor is reconstructed using each sub-dictionary individually and every test B-scan is assigned to the class with the minimal residual error. To assess the generalizability of the model, we have tested it on three different databases. Furthermore, we introduce a new heatmap generation approach based on averaging the most significant atoms of the learned sub-dictionaries, demonstrating that selecting an appropriate sub-dictionary for test B-scan restoration can lead to better reconstructions, emphasizing distinctive features of different classes. CircWaveDL demonstrates a high level of generalizability according to external validation conducted on three different databases and it outperforms previous classification methods designed for similar datasets.