Hybrid Method for Ultrasound Image Enhancement Based on Deep Learning Technology

Abstract

Noise removal is an essential preprocessing step, especially with medical images. When dealing with ultrasound images, it is extremely important to ensure the success and accuracy of the diagnosis. Due to some factors accompanying the image capture process, we notice some noise accompanying the image, which leads to difficulty in the process of seeing fine and sensitive tissues and complex areas correctly. There are many techniques used in this field, which are well-known but have certain limitation. On the other hand, there are some techniques such as deep networks, which have proven their worth, but they consume large amounts of data in the training phase.

In this research, a new method is presented that combines a Residual Dense Network (RDN)   with adaptive noise estimation techniques, some traditional filters, and deep learning techniques. The goal of this combination is to improve the clarity and contrast of radiological images. The rationale behind this approach is that the two methods will compensate for each other's weaknesses while reinforcing their respective strengths. Experiments on a model showed great improvement in the quality of the images, as validated by the PSNR, SSIM, NIQUE index, direct rate, mean square error rate. The model was tested over a set of large radiological images in benign as well as in malignant cases, and the noise effect was investigated in terms of 20, 50, 80, and 100 levels.