Simultaneous reconstruction and restoration of sparsely sampled optical coherence tomography image through learning separable filters for deep architectures

Spectral domain optical coherence tomography (SD-OCT) is widely employed across ophthalmology practices for visual investigation of live tissues. The involuntary movements of subjects frequently infuse motion artifacts to SD-OCT images. Sub-sampling of signals is introduced in imaging protocol to avoid such artifacts which causes fall in spatial resolution and peak signal to noise ratio (PSNR). Sparse coding (SC) is opted for restoration and rectification of complete signals from sparse samples through constructing complete and sparse space dictionaries independently. Convolutional neural networks (CNN) can be casted as SC for jointly learning dictionaries resulting less number of CNN filters (equivalence of SC dictionaries) to be trained. The proposed approach extends the separable filters to CNN through architectural constrain. This results in a parallel architecture and reduced number of parameters without compromising on performance. The approach scaled down trainable parameters by 46% with a trade-off of 0.108 PSNR during training and 0.107 PSNR during testing in comparison to conventional CNN.