Despeckling filters applied to thyroid ultrasound images: a comparative analysis

The speckle noise is an intrinsic artefact present in ultrasound images that masks the diagnostically important information, thus makes it hard for the radiologists to analyze them. Therefore, a suitable despeckling algorithm, which will retain the diagnostically important features such as structure, edges and margins are required. In this study the performance of 64 despeckling filters algorithms used for the analysis of thyroid nodule ultrasound images is compared. These 64 filters are divided into 9 categories namely Linear, Non-linear, Total Variation, Fuzzy, Fourier, Multiscale, Nonlocal Mean, Edge Preserving, and Hybrid filters. A total of 820 thyroid US images have been taken from two different benchmark datasets. Out of these 820 thyroid US images, 200 are benign and 620 are malignant. The performance analysis of despeckling filters has been carried out by calculating structure and edge preservation index metric. It has been observed that fast bilateral filter and edge-preserving smoothing filter yields optimal performance with respect to the preservation of image structures like edges and margins of benign and malignant thyroid tumors. Based on the criterion followed in real time clinical practice for differential diagnosis between benign and malignant thyroid ultrasound tumors, it is observed that the images filtered by DsF_EPSF filter yields better diagnostic quality images in terms of preservation and enhancement of important diagnostic information.