Ultrasound despeckling by anisotropic diffusion and total variation methods for liver fibrosis diagnostics

Anisotropic diffusion method and a total variation method for B-mode ultrasound image speckle filtering were compared for the problem of liver fibrosis diagnostics. An effective stopping criterion to control the strength of image filtering with the anisotropic diffusion algorithm and the regularization parameter estimation method for the proposed total variation algorithm were introduced. The comparison of two speckle filtering techniques demonstrated the advantage of anisotropic diffusion algorithm. Liver fibrosis diagnostics was performed using image texture analysis based on 10-20 textural characteristics. Siemens ACUSON S2000 ultrasound images of liver for 60 patients were used to determine the fibrosis according to the METAVIR score. A two-step algorithm includes elastography based F4 stage detection and F0–(F1,F2,F3) separation using ultrasound texture analysis. The classification was performed with Random Forest classifier. A comparison with deep convolutional neural networks was also performed. It was found that speckle filtering procedure in some cases enhances the texture-based classification and increases the total accuracy value up to 5% and makes the classification more robust and independent from the train–test set selection. •Anisotropic diffusion and total variation methods for image despeckling are compared.•A two-step algorithm for fibrosis stage determination is proposed.•The algorithm includes the elastography and ultrasound image texture analysis.•The total classification accuracy in case of separation into 3 classes is 90%.•Despeckling makes the classification more independent from train–test set selection.