Innovative temporal loss function for segmentation of fine structures in ultrasound images

Over the past few years, there have been significant advancements in deep learning architectures for semantic segmentation. However, the performance of these models heavily relies on the loss function (LF) used during network training. The LF is a crucial component that enables the network to weigh the errors differently based on the segmentation task to be performed. Despite the progress made in designing increasingly complex and deep architectures for semantic segmentation, the LFs used in these models have remained almost unchanged. Accurately segmenting small and fine objects, such as vessel walls (e.g., intima-media complex, IMC) or nerves (e.g., optic nerve), in ultrasound (US) images is still a challenging task. One of the main difficulties is pixel imbalance between the object and the background, which can result in inaccurate segmentation. Additionally, precise and accurate segmentation along the object's edge is crucial for medical diagnosis and treatment. To address these challenges, this paper proposes a new, temporal loss function for semantic segmentation in US images. The idea behind a temporal loss is to enable the network to learn from multiple sources of information simultaneously and to give more emphasis to losses that are more informative at different stages of the training process. The proposed LF considers pixel imbalance between the object and background and enables precise and accurate segmentation along the object's edge. The study aims to demonstrate the effectiveness of the proposed LF by evaluating its performance in segmenting vessel walls in US images.