Computer-aided systems can help the ophthalmologists in early detection of most of ocular abnormalities using retinal OCT images. The need for more accurate diagnosis increases the need for modifications and innovations to current algorithms. In this paper, we investigate the effect of different X-lets on the classification of OCT B-scans of a dataset with one normal class and two abnormal classes. Different transforms of each B-scan have been fed to the designed 2D-Convolutional-Neural-Network (2D-CNN) to extract the best-suited features. We compare the performance of them with MSVM and MLP classifiers. Comparison with the accuracy of normal and abnormal classes reveals substantially better results for normal cases using 2D-Discrete-Wavelet-Transform (2D-DWT), since the structure of most normal B-scans follows a pattern with zero-degree lines, while for abnormalities with circles appearing in the retinal structure (due to the accumulation of fluid), the circlet transform performs much better. Therefore, we combine these two X-lets and propose a new transform named CircWave which uses all sub-bands of both transformations in the form of a multi-channel-matrix, with the aim to increase the classification accuracy of normal and abnormal cases, simultaneously. We show that the classification results obtained based on CircWave transform outperform those based on the original images and each individual transform. Furthermore, the Grad-CAM class activation visualization for B-scans reconstructed from half of the CircWave sub-bands indicates a greater focus on appearing circles in abnormal cases and straight lines in normal cases at the same time, while for original B-scans the focus of the heat-map is on some irrelevant regions. To investigate the generalizability of our proposed method we have applied it also to another dataset. Using the CircWave transform, we have obtained an accuracy of 94.5% and 90% for the first and second dataset, respectively, while these values were 88% and 83% using the original images. The proposed CNN based on CircWave provides not only superior evaluation parameter values but also better interpretable results with more focus on features that are important for ophthalmologists.
