Abstract
Purpose: :
Post acquisition image processing is frequently used to align and denoise Spectral Domain Optical Coherence Tomography (SD-OCT) data. A standardized image database would facilitate fair and quantitative comparisons of the effects of image processing algorithms. The purpose of this study was to develop a simplified statistical model representing the main structural layers of the retina and to compare the performance of two popular algorithms for image denoising based on wavelets and curvelets.
Methods: :
The model developed includes 7 homogeneous regions representing distinct layers of the retina in the Brown Norway rat. An expert manually segmented each retinal layer from SD-OCT B-scans and statistical parameters for the distribution of pixel intensity were derived for each layer. Using this model, ten synthetic SD-OCT images (512x512 px) were generated and speckle noise was added to obtain a set of images in which the signal-to-noise ratio (SNR) ranged between 35 and 55 dB. These images were then filtered using stationary wavelets (Symlets) and curvelets to obtain estimates of the original noise-free synthetic images using standard deviation based soft thresholding schemes. The performance of the two denoising methods were assessed using the mean square error (MSE), power of MSE image, and power of the absolute difference image before and after filtering.
Results: :
Using the MSE measure, the average performance of curvelets was 52.4% ± 4.57% better compared to wavelets in the low SNR range (35 to 45 dB) and 60.76% ± 3.65% in the high SNR range (45 to 55 dB). Using the power of MSE image measure, the average performance of curvelets was 57% ± 19% better compared to wavelets in the low SNR range and the average performance of wavelets was 58.3% ± 30.9% better compared to curvelets in the high SNR range. Using the power of absolute difference image measure, the average performance of curvelets was 32.78% ± 14.34% better compared to wavelets in the low SNR range and the average performance of wavelets was 34% ± 17.38% better compared to curvelets in the high SNR range.
Conclusions: :
Our results show that curvelet denoising outperformed stationary wavelet denoising over a wide range of SNRs. Furthermore, the proposed model used to generate B-Scan images not only facilitated the quantitative comparison of image denoising algorithms, but it can also be an invaluable tool in assessing the performance for image segmentation and feature extraction algorithms.
Keywords: image processing • imaging/image analysis: non-clinical • computational modeling