Echocardiography (echo) is a non-invasive, safe, widely available imaging modality that is frequently used to assess the heart structure and function. Accurate heart chamber segmentation is an essential step to quantify certain parameters, including heart chamber volumes. In clinical practice, this task is manually done by echo experts, where it consumes considerable time and is subjective to both errors as well as intra-operator variability. Artificial intelligence (AI) models have been used to automatically segment heart chambers. We conducted a scoping review to provide an overview of the AI models used for this task. Three bibliographic databases; PubMed, Embase, and Google Scholar were explored. Out of 640 initially retrieved studies, 36 studies were included. Multiple AI models used for echo images segmentation were identified, which can be broadly categorized into five methods: low-level image processing, deformable-based, statistical techniques, machine learning (ML), and deep learning-based (DL) techniques. The initial three categories were relatively simple and required less computational complexity compared to the ML and DL models. The convolutional neural network was the most widely used DL-based technique in most-recent publications. Generalizability of the models is a major concern that needs to be addressed in the future. Well-annotated larger 2D echo image datasets would be required to mitigate the challenges to some extent.
Keywords: 2D echocardiography; Artificial intelligence; Deep learning; segmentation.