Attention Based Speckle Reduction for Optical Coherence Tomography in Ophthalmology and Dentistry

An emerging imaging tool in medical imaging is Optical Coherence Tomography (OCT) that is widely implemented in multiple fields such as ophthalmology for detection of retinal diseases. In this paper a gated-filtering deep learning denoising framework is proposed for OCT images, that presents a new multi-kernel filtering block integrated in CNN and incorporates the attention mechanism. The main usage of attention gates is to ensure focus on speckle removal of the foreground and to eliminate the background. The filtering block is designed to remove the different types of noise artifacts, this is shown visually and compared against well-known denoising methods. Extensive quantitative evaluations are performed, in which the proposed methodology displays quantitively verifiably results that removes speckle noise and achieves leading image quality to state-of-the-art denoisers. The proposed method improved the PSNR by 26.7 dB, CNR by 7.2 dB and ENL by 588.5 dB for retinal dataset and by 26.9 dB, 7.0 dB and 213.7 dB respectively for a dentistry dataset compared well-known denoisers.