A Novel Improved Framework for Speckle Denoising using Deep Learning Approach

Ultrasound technology serves as a popular non-invasive diagnostic tool. Speckle noise (SN), a common artifact in ultrasound images can diminish their resolution and contrast potentially affecting the accuracy of disease diagnosis. Researchers have been focused on sophisticated Deep Learning (DL) models to remove simulated additional SN they have not been concentrating on the practical use of DL to remove naturally occurring SN from genuine US images. Mostly DL algorithms used Autoencoder models and CNN based techniques to do so but they experience overfitting, decreased accuracy, and performance deterioration, making it difficult to give enhanced output of ultrasound images. Inspired by the researchers, we proposed Unet-Elu CNN model with Exponential activation function (Elu function). The improved Unet-ELU model consists of 26 layers. Batch normalization, max pooling layers and drop out layers are added for more accuracy and preventing overfitting issue. The objective of this proposed work is to improve the resolution and minimize speckle noise in medical ultrasound images to make images more clear and enhanced for diagnosis of disease. Model has been trained using breast cancer and abdominal dataset, the original US images were given as targets, while the noised US images served as the input. This model was evaluated using the evaluation metrics of PSNR, SSIM, and MSE with different levels of speckle noise in order to determine the effectiveness of these techniques. In comparison with Unet, Leaky relu, CNN Auto-encoder model, and other filtering techniques, proposed framework obtained a PSNR value of 37.76 and SSIM 98%, respectively, indicating better results.