Noise reduction by adaptive-SIN filtering for retinal OCT images

Yan Hu, Jianfeng Ren, Jianlong Yang, Ruibing Bai, Jiang LiuScientific 2021 · DOI

TL;DR

This paper introduces an "adaptive-SIN" algorithm to significantly improve the clarity of retinal Optical Coherence Tomography (OCT) images by reducing noise. It achieves this by first transforming the signal's inherent Poisson noise into Gaussian noise, then applying a 3D shearlet transform for decomposition, and finally using an adaptive thresholding method to effectively remove noise while preserving crucial image details.

What Problem Does This Paper Address?

Retinal OCT imaging is a vital tool for ophthalmic examinations, providing high-resolution, 3D views of the eye's internal structures. However, these images are often plagued by noise, which can obscure important details, hinder accurate diagnosis, and affect the precision of surgical instruments. Existing denoising methods, such as averaging or standard filter-based approaches, have limitations. Some struggle with the specific noise characteristics of OCT (which are closer to Poisson than Gaussian), while others fail to adapt to variations in noise levels caused by different OCT machine manufacturers, imaging protocols, or even the specific biological tissues being imaged. This variability makes a one-size-fits-all denoising approach ineffective.

Methods

The proposed "adaptive-SIN" (adaptive Shearlet Image-filtering with Noise-redistribution) algorithm tackles these challenges through a multi-step process:

  1. Noise Redistribution (Square-Root Transform): Recognizing that OCT noise is closer to a Poisson distribution and shearlet transforms typically assume Gaussian noise, the authors introduce a square-root transform. This mathematical step effectively converts the Poisson noise into a distribution that better approximates Gaussian noise, preparing it for the subsequent shearlet filtering.
  2. 3D Shearlet Decomposition: Instead of traditional 2D methods, the algorithm employs a 3D shearlet transform. This is crucial because retinal OCT data is inherently volumetric (multiple B-scans forming a 3D volume), and the 3D shearlet transform can better capture inter-frame information and directional sensitivities within this volume, leading to more effective noise removal and detail preservation.
  3. Adaptive Thresholding: A key innovation is the adaptive thresholding scheme. Unlike fixed thresholds that struggle with varying noise conditions, this method dynamically adjusts the threshold based on the estimated noise variance in different subbands of the shearlet decomposition. This allows the algorithm to intelligently decide how much noise to remove in each region, preventing over-smoothing or under-denoising.
  4. Shearlet Reconstruction: Finally, an inverse 3D shearlet transform reconstructs the denoised image from the processed coefficients, yielding a high-quality OCT image.

Datasets

The proposed adaptive-SIN algorithm was evaluated on three distinct datasets:

  • Normal Subject Eyes Dataset: This dataset consists of 20 3D retinal OCT volumes from 20 healthy subjects, acquired using a Topcon ATLANTIS OCT machine. For comparison, "ground-truth" noiseless images were generated using a line-scan mode, which repeatedly scans the same position to average out noise.
  • Patients Dataset (Topcon OCT Machine): This dataset includes 10 3D volumes from patients' eyes, captured with a Topcon ATLANTIS OCT machine. Due to the clinical setting, ground-truth noiseless images were not available for this dataset.
  • Patients Dataset (Optovue OCT Machine): This dataset comprises 10 3D volumes from patients' eyes, acquired using an Optovue OCT machine, which has different lateral and in-depth resolutions compared to the Topcon machine. Ground-truth noiseless images were also unavailable for this dataset.

Key Insights and Ideas

  • Poisson to Gaussian Noise Transformation is Crucial: The paper highlights that OCT noise isn't always Gaussian. Transforming it to a Gaussian-like distribution via a simple square-root function is a critical preprocessing step that significantly enhances the effectiveness of shearlet-based denoising.
  • 3D Shearlets for Volumetric Data: Moving from 2D to 3D shearlet transforms is essential for leveraging the volumetric nature of OCT data, allowing for better capture of inter-frame correlations and directional features, which are vital for accurate denoising.
  • Adaptability is Key for Real-World OCT: The core insight is that noise characteristics vary widely across different OCT machines and imaging scenarios. An adaptive thresholding mechanism is not just beneficial but necessary to achieve robust and effective denoising across diverse datasets.
  • Balancing Noise Removal and Detail Preservation: The adaptive-SIN algorithm demonstrates a superior ability to strike a balance between aggressively removing noise and preserving fine image structures like edges and textures, which are critical for clinical diagnosis.

Key Results

The adaptive-SIN algorithm demonstrated superior performance across all tested datasets and evaluation metrics.

  • Quantitative Improvements: The paper reports significant improvements in metrics like Peak Signal-to-Noise Ratio (PSNR), Mean Structural Similarity Index (MSSIM), and Edge Preservation Index (EPI) compared to various state-of-the-art methods, including 2D and 3D shearlet filters, ASR, SBD, kSVD, BM4D, cGAN, and ReShearlet. For instance, in Table 2, the proposed method achieved the highest PSNR (22.51) and MSSIM (0.78), along with the lowest MSE (0.057) and highest EPI (0.88) and ENL (19.46).
  • Qualitative Improvements: Visual inspection of the denoised images consistently showed that the adaptive-SIN algorithm produced cleaner images with significantly less noise while retaining finer details and sharper edges compared to other methods. Figures 6, 7, 8, and 9 visually confirm these findings, showcasing the algorithm's ability to produce high-quality OCT images suitable for clinical analysis.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, multi-dimensional shearlet transforms and adaptive processing can be computationally intensive. The paper mentions experiments were conducted on a dual-core 2.20GHz CPU, suggesting that optimization for real-time applications might be a future consideration.
  • Ground Truth Availability: For patient datasets from Topcon and Optovue machines, the absence of ground-truth noiseless images meant that evaluation relied on metrics like EPI, ENL, and CNR, which are less direct measures of overall image fidelity compared to PSNR and MSE.
  • Future Work: The authors suggest that the denoised images can be directly used for OCT layer segmentation, which is a foundational step for various retinal disease analyses and correlation studies. This implies that further integration of the denoising technique into downstream clinical analysis pipelines is a natural next step.

Summary updated

Abstract

Optical coherence tomography (OCT) images is widely used in ophthalmic examination, but their qualities are often affected by noises. Shearlet transform has shown its effectiveness in removing image noises because of its edge-preserving property and directional sensitivity. In the paper, we propose an adaptive denoising algorithm for OCT images. The OCT noise is closer to the Poisson distribution than the Gaussian distribution, and shearlet transform assumes additive white Gaussian noise. We hence propose a square-root transform to redistribute the OCT noise. Different manufacturers and differences between imaging objects may influence the observed noise characteristics, which make predefined thresholding scheme ineffective. We propose an adaptive 3D shearlet image filter with noise-redistribution (adaptive-SIN) scheme for OCT images. The proposed adaptive-SIN is evaluated on three benchmark datasets using quantitative evaluation metrics and subjective visual inspection. Compared with other algorithms, the proposed algorithm better removes noise in OCT images and better preserves image details, significantly outperforming in terms of both quantitative evaluation and visual inspection. The proposed algorithm effectively transforms the Poisson noise to Gaussian noise so that the subsequent shearlet transform could optimally remove the noise. The proposed adaptive thresholding scheme optimally adapts to various noise conditions and hence better remove the noise. The comparison experimental results on three benchmark datasets against 8 compared algorithms demonstrate the effectiveness of the proposed approach in removing OCT noise.
BibTeX
@inproceedings{hu2021noise,
  author = {Yan Hu and Jianfeng Ren and Jianlong Yang and Ruibing Bai and Jiang Liu},
  title = {{Noise reduction by adaptive-SIN filtering for retinal OCT images}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-98832-w},
}
Figure