Comparison of Noise Reduction Algorithms for Optical Coherence Tomography Images of Skin Melanoma

• Published in: Izvestiâ vysših učebnyh zavedenij Rossii. Radioèlektronika Vol. 23; no. 4; pp. 66 - 76
• Main Author: Myakinin, O. O.
• Format: Journal Article
• Language: English; Russian
• Published: Saint Petersburg Electrotechnical University "LETI" 07.10.2020
• Subjects: винеровский фильтр; диффузионный фильтр; медианный фильтр; нечеткая логика; оптическая когерентная томография; эмпирическая модовая декомпозиция
• ISSN: 1993-8985; 2658-4794
• DOI: 10.32603/1993-8985-2020-23-4-66-76

Introduction. Optical coherence tomography (ОКТ) is a non-invasive instrument for studying optically heterogeneous media with micron precision, including skin cancer. However, ОКТ tissue images are very noisy. It complicates both expert and automated image evaluations. There are almost no systematic comparisons of noise reduction algorithms in the literature.Objective. To obtain comparative test results on a set of ОКТ images of skin melanoma using various noise reduction algorithms.Materials and methods. A number of noise reduction algorithms were described, which include two relatively simple classical algorithms: Wiener and median, and more complex ones: a Complex Diffusion Filter (CDF), an Interval type-II Fuzzy Anisotropic Diffusion Filter (ITTFADF) and an Empirical Mode Decomposition (EMD) filter, previously proposed by the author for visualizing of mesh implants. Quantitative metrics were determined: a Signal-to-Noise Ratio (SNR) metrics, an Effective Number of Looks (ENL) metrics, Structural Similarity Index Metrics (SSIM) and a correlation coefficient χ, reflecting two main principles of improving image quality: to reduce noise and to save the borders of tissue layers and heterogeneities.Results. The results of a comparative testing on a set of images, consisting of 10 melanomas (to which various noise reduction algorithms were applied) were obtained.Conclusion. The study did not reveal the best algorithm for all four metrics. According to the SNR metric, the EMD and the CDF filters perform the best depending on the type of area. At the same time, the EMD filter is either the best in all respects, or is inferior in SNR in heterogeneous areas and takes the second place in ENL. Taking as the correct hypothesis that the border preservation is more important before an integral noise estimate, it is possible to make an unambiguous conclusion about the need to use the EMD filter. As an alternative to the EMD filter, Wiener filter (which wins on the border preservation metrics) should be used or the ITTFADF, which ranked third in all used metrics.