Enhanced U-net architecture for ultrasound image speckle reduction

In the modern world, although ultrasound imaging is the most popular medical imaging technology and is mostly utilized for illness diagnosis, interference and low image intensity cause the margins and boundaries of the images to be less distinct than they should be. In ultrasonic images, speckle noise is the primary cause of image quality loss. For radiologists to have better images to aid in their diagnosis process, this noise must be eliminated. The proposed model aim is to enhance the resolution and reduce speckle noise in medical ultrasound images using deep learning approach. The proposed Unet_elu model is a CNN model based on U-net architecture with Exponential activation functions (Elu function) got better results than the Autoencoder model. This model has been trained using breast cancer dataset and Compared the results with autoencoder model and various filtering methods like Median blur, Gaussian blur, Average blur and Bilateral filtering. Google Colaboratory, often known as Colab, is a Python language program used for the simulation. The resulting reconstructed image was assessed using the evaluation metrics of peak signal to noise ratio (PSNR), structural similarity index measure (SSIM), and mean squared error (MSE), with varying degrees of speckle noise. Based on the obtained results, the research has determined that proposed algorithm has provided better results for reducing speckle noise and producing enhanced output images compared to CNN Autoencoder model and other filtering methods. We got a PSNR value 32.38 and SSIM 95% which is better than the other filtering methods and traditional methods of image enhancement. Physicians and radiologists can use the suggested Unet_elu models to help in validating their initial screening for ultrasound images.