Evaluation of Denoising Performance of ResNet Deep Learning Model for Ultrasound Images Corresponding to Two Frequency Parameters

Abstract

Ultrasound imaging is widely used for accurate diagnosis due to its noninvasive nature and the absence of radiation exposure, which is achieved by controlling the scan frequency. In addition, Gaussian and speckle noises degrade image quality. To address this issue, filtering techniques are typically used in the spatial domain. Recently, deep learning models have been increasingly applied in the field of medical imaging. In this study, we evaluated the effectiveness of a convolutional neural network-based residual network (ResNet) deep learning model for noise reduction when Gaussian and speckle noises were present. We compared the results with those obtained from conventional filtering techniques. A dataset of 500 images was prepared, and Gaussian and speckle noises were added to create noisy input images. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The ResNet deep learning model, comprising 16 residual blocks, was trained using optimized hyperparameters, including the learning rate, optimization function, and loss function. For quantitative analysis, we calculated the normalized noise power spectrum, peak signal-to-noise ratio, and root mean square error. Our findings showed that the ResNet deep learning model exhibited superior noise reduction performance to median, Wiener, and median-modified Wiener filter algorithms.

Keywords: deep learning; frequency; quantitative analysis; residual network (ResNet); ultrasound.

Figure 1

Experimental setup including acquisition of dataset at 3 and 5 MHz.

Figure 2

The ResNet model architecture with 16 residual blocks.

Figure 3

Resulting in images for 3 MHz with (a) 0.01 and (b) 0.1 standard deviation values of Gaussian and speckle noise distribution corresponding to the noisy image, median filter, Wiener filter, MMWF algorithm, and proposed model, respectively.

Figure 4

Resulting in images for 5 MHz with (a) 0.01 and (b) 0.1 standard deviation values of Gaussian and speckle noise distribution corresponding to the noisy image, median filter, Wiener filter, MMWF algorithm, and proposed model, respectively.

Figure 5

Normalized noise power spectrum results for spatial resolution with (a) 0.01 and (b) 0.1 of standard deviation values of Gaussian and speckle noise distribution corresponding to noisy, median filter, Wiener filter, MMWF algorithm, and proposed model for 3 MHz.

Figure 6

Normalized noise power spectrum results for spatial resolution with (a) 0.01 and (b) 0.1 of standard deviation values for Gaussian and speckle noise distribution corresponding to noisy, median filter, Wiener filter, MMWF algorithm, and proposed model for 5 MHz.

Figure 7

Similarity results between label and output images for 3 MHz for peak signal-to-noise ratio and root mean squared error corresponding to (a,b) 0.01 and (c,d) 0.1 of standard deviation values of Gaussian and speckle noise.

Figure 8

Similarity results between label and output images for 5 MHz for peak signal-to-noise ratio and root mean squared error corresponding to (a,b) 0.01 and (c,d) 0.1 standard deviation values of Gaussian and speckle noise.

Figure 9

Intensity profile including magnified portion according to (a) 0.01 and (b) 0.1 standard deviation values of Gaussian and speckle noise at 3 MHz.

Figure 10

Intensity profile including magnified portion according to (a) 0.01 and (b) 0.1 standard deviation values of Gaussian and speckle noise at 5 MHz.