Enhanced Denoising of Optical Coherence Tomography Images Using Residual U-Net
Authors:
Akkidas Noel Prakash,
Jahnvi Sai Ganta,
Ramaswami Krishnadas,
Tin A. Tunc,
Satish K Panda
Abstract:
Optical Coherence Tomography (OCT) imaging is pivotal in diagnosing ophthalmic conditions by providing detailed cross-sectional images of the anterior and posterior segments of the eye. Nonetheless, speckle noise and other imaging artifacts inherent to OCT impede the accuracy of diagnosis significantly. In this study, we proposed an enhanced denoising model using a Residual U-Net architecture that…
▽ More
Optical Coherence Tomography (OCT) imaging is pivotal in diagnosing ophthalmic conditions by providing detailed cross-sectional images of the anterior and posterior segments of the eye. Nonetheless, speckle noise and other imaging artifacts inherent to OCT impede the accuracy of diagnosis significantly. In this study, we proposed an enhanced denoising model using a Residual U-Net architecture that effectively diminishes noise and improves image clarity across both Anterior Segment OCT (ASOCT) and polarization-sensitive OCT (PSOCT) images. Our approach demonstrated substantial improvements in image quality metrics: the Peak Signal Noise Ratio (PSNR) was 34.343 $\pm$ 1.113 for PSOCT images, and Structural Similarity Index Measure (SSIM) values were 0.885 $\pm$ 0.030, indicating enhanced preservation of tissue integrity and textural details. For ASOCT images, we observed the PSNR to be 23.525 $\pm$ 0.872 dB and SSIM 0.407 $\pm$ 0.044, reflecting significant enhancements in visual quality and structural accuracy. These metrics substantiate the models efficacy in not only reducing noise but also in maintaining crucial anatomical features, thereby enabling more precise and efficient clinical evaluations. The dual functionality across both ASOCT and PSOCT modalities underscores the versatility and potential for broad application in clinical settings, optimizing diagnostic processes and reducing the necessity for prolonged imaging sessions.
△ Less
Submitted 24 September, 2024; v1 submitted 17 July, 2024;
originally announced July 2024.
