Effect of Speckle Filtering in the Performance of Segmentation of Ultrasound Images Using CNNs

Abstract

The convolutional neural networks (CNNs) as tools for ultrasound image segmentation often have their performance affected by the low signal-to-noise ratio of the images. This prevents a correct classification and extraction of relevant information and therefore affects clinical diagnosis. We propose a study of the effect of different speckle filtering methods on CNN performance. For the proposed metrics (Jaccard coefficient and BF-Score), it was obtained that the SRAD filter exhibited the best behavior even in the lowest quality data. In addition, the lowest values were obtained for the standard deviation and variance, which translates into lower data dispersion, better repeatability, and, therefore, greater confidence in its accuracy. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.