A Lightweight Mimic Convolutional Auto-Encoder for Denoising Retinal Optical Coherence Tomography Images

Optical coherence tomography (OCT) is widely used for diagnosing and monitoring retinal disorders. However, despite hardware improvements, its scans are still highly affected by speckle noise. Speckle noise reduces quality of measurements and decreases reliability of further instrumentation. Recent OCT denoising methods are often complex and computationally inefficient, despite their valid performance. These methods can be used as reference methods to train deep auto-encoders (AEs). AE networks can learn important structural features of OCT images that have been denoised with these reference methods and use features to reconstruct or denoise corrupted ones. In this way, a well-trained AE can efficiently mimic that reference denoising method. In this study, we implemented a lightweight convolutional AE to mimic a recent state-of-the-art method in OCT image denoising. We evaluated the performance of AE for various test data sets using both visual inspection and quantitative metrics. Presented results confirmed good performance of the proposed AE in despeckling OCT scans. Results revealed the generality, computationally efficiency, and device independence property of the proposed method. These features make the proposed network applicable in real time, mobile application due to its high denoising speed and low memory usage.

Automated summary

The study implemented a lightweight convolutional AE network for mimicking the latest OCT image denoising method. The performance of the network was evaluated on various test data sets using visual inspection and quantitative metrics, confirming its good performance in removing speckle noise in OCT scans. The proposed network demonstrated generality, computational efficiency, and device independence, making it suitable for real-time, mobile applications.