چكيده به لاتين
Cardiovascular diseases (CVDs) are the main cause of mortality around the world. The first step for the control and treatment of these diseases is an accurate diagnosis, which is achievable by diagnostic tools such as imaging systems (e.g. echocardiography, angiography, magnetic resonance imaging, etc.). Among these imaging systems, echocardiography undoubtedly is the preferred device for the evaluation of cardiac ventricles. Although cardiac magnetic resonance imaging (CMRI) provides a better visualization of anatomical structures and wall motion, echocardiography is more popular due to its low cost, temporal resolution, and portability. To have an efficient diagnosis, there is a crucial need for left ventricle (LV) segmentation to be used in calculations of clinical indices such as end diastolic (ED) and end systolic (ES) volumes, ejection fraction, left ventricular mass, etc.
In this research, our goal is the automatic segmentation of cardiac echocardiography images using some methods based on deep learning on the largest publicly-available and fully-annotated dataset for the purpose of echocardiographic assessment (CAMUS). The dataset contains two and four-chamber acquisitions from 500 patients with reference measurements from one cardiologist on the full dataset and from three cardiologists on a fold of 50 patients. The methods used in this review are the improved methods of the U-net network.
In this research, 6 deep learning models have been used to segment the left ventricle of cardiac echocardiography images presented in the CAMUS dataset. 3 of these models named U-Net 1, U-Net 2 and MFP-UNet were the reference models presented and tested on the mentioned dataset. The Modified U-net network has been the proposed model for the segmentation of head and neck tumors, which was used to segment the left ventricle in this data set by changing the number of convolution filters in each layer. The MFP-Unet 1 model is an improved model of the MFP-Unet network that is used in this research and at the end, the model named AW-Net is presented by the author, which is designed based on the attention mechanism. Also, in this research, 3 strategies have been used to train networks. The difference between these strategies is in the input of the networks. The first strategy is that the input of the networks is raw echocardiography images without applying any filters. The second strategy is that the input of networks are filtered images with two SRAD and intermediate filters. The third strategy, which is the input of the appended models, is two raw and filtered images that enter the network as a 2-channel input.
Among the proposed models in terms of Dice metric and Hausdorff distance, the AW-Net model shows the best performance, the Dice metric of this model is 0.941 ± 0.002 and its Hausdorff distance is 2.94 ± 0.06 mm. Also, this model is in a very good position in terms of the balance between network simplicity, network speed, and the number of learnable parameters with the proposed geometric metrics. After training these networks in all three strategies, it was seen that the results of the networks in the second strategy were better than the results of the first strategy, but they did not differ much in the third strategy, even in some models it was seen that the results were worse than the first strategy. Dice metric of AW-Net model in the second strategy is equal to 0.943 ± 0.002 and its Hausdorff distance is equal to 2.93 ± 0.05 mm.
