accurate left ventricular segmentation in echocardiographic using unet architecture

Introduction: heart disease has been the leading cause of death worldwide since 1991, with one person dying from the condition every 1.5 seconds [1]. this worrying prevalence highlights the urgent need for rapid and precise diagnostic methods, which remain a significant challenge.echocardiography serves as a crucial solution, offering a portable, non-invasive, and cost-effective imaging technique for the initial diagnosis of heart diseases. this technique provides the ability to identify clinical parameters, including left ventricular volume and ejection fraction.manual segmentation by cardiologist is often time-consuming, challenging, and tedious work. while conventional algorithms try to compensate for these challenges, they are still facing high error rates. on the other hands, advances in artificial intelligence have significantly mitigated these obstacles [2].methods and materials: in this study, we utilized the camus dataset published by leclerc et al., which includes 2d sequences of two and four-chamber views from 450 patients, featuring images of varying quality. images were captured at both end-diastole (ed) and end-systole (es), accompanied by ground truth annotations of the left ventricle [3]. to validate the algorithm's performance, we tested it on data from 30 patients, annotated by an expert cardiologist, which included two-dimensional sequences of both two and four-chamber views, as well as a three-chamber view.deep learning networks (unet) are utilized to segment the left ventricle, as they are particularly effective because of two important parts: the encoder and the decoder. in the encoder path, the input data is processed using many layers of convolution and down-sampling. this sequence improves feature abstraction, making higher-level representations. in contrast, the decoder path uses a series of up-sampling stages to restore the original image's resolution. also, the connection layer, which connects the encoder and decoder, results in more accurate segmentation results [4].results: a deep learning model was implemented in python using the pytorch library on the google colab platform. the dataset included 450 patients from the camus study, of which 80% were used for training and the remaining 20% for evaluation. we used our own labeled data, which included 30 patients, to test the network (fig.1).to evaluate the accuracy of the segmentation results, we applied the dice similarity coefficient (dsc). this statistical measure determines the overlap between the predicted segmentation and the ground truth. this measure is particularly sensitive to precisely predicted values. on the evaluation dataset, the unet model achieved a dsc value of 91.6%. meanwhile, the test dataset had a dsc of 77.9%, showing a slight reduction in segmentation accuracy.conclusion and discussion: the unet architecture demonstrated promising results for accurate segmentation of the left ventricle in echocardiographic images. notably, it performed well despite not being trained on our collected data, which also includes three-chamber views. this method has the potential to serve as a valuable clinical tool for assessing heart function. future studies should explore the application of this approach on larger and more diverse datasets to further validate its effectiveness.