Impacts of Losses Functions on the Quality of the Ultrasound Image by Using Machine Learning Algorithms

Abstract

During last decade, Artificial Intelligence (AI) has been able to reshape our life daily. Different areas were positively impacted by AI such as Healthcare, Logistic, etc. Medical imaging is one of the fields of healthcare in which AI was introduced to solve and overcome different problems. Challenges including image processing, signal processing, and data acquisition. In this paper, we deeply demonstrate the loss function as one of the main parameters that influence the quality of the ultrasound (US) image. Therefore, we introduce the main components of ultrasound systems form end-to-end perspective such as the data acquisition, the signal processing, and the image interpretation. Then, we present the losses functions as a critical performance metrics for the model validation. Metrics such as the Mean Absolute Error (MAE), Cross-Entropy loss function (CE), Dice Similarity Coefficient (DSC), and the Structural Similarity (SSIM). After that we present the adopted CNN model to generate ultrasound image. The excessive simulation results demonstrate that the selection of the loss function provides significant improvement in terms of image quality (e.g., contrast, CNR and SNR). Choosing simple loss functions such as mean square error helps to faster the convergence of the convolution neural network during the training process. However, for image quality enhancement, we propose the combination of different loss functions such structural similarity (SSIM) with Dice Similarity Coefficient (DSC).