سعيد چمني

Automatic segmentation of 2D echocardiograms is a critical step in diagnosing an‎d assessing heart diseases. Convolutional Neural Networks (CNNs), such as U-Net, have proven effective for medical image segmentation, but they often fall short in capturing long-range dependencies because convolution operations are inherently local. While transformer models capture the global context, they also may not be good in fine-grained localization. To han‎dle these issues, we introduce a novel deep learning architecture based on an encoder-decoder framework, enriched by multiple attention mechanisms, an‎d hybrid bottleneck combining a Vision Transformer (ViT) an‎d a Multi Receptive Field Block (MRFB). The integration of a ViT enables the modeling of long-range dependencies, while the MRFB improves multi-scale feature extraction an‎d is specifically designed to improve the representation of spatial features an‎d reduce the loss of spatial information. Enhanced by multiple attention mechanisms an‎d deep supervision with Atrous Spatial Pyramid Pooling (ASPP), our model robustly captures both local an‎d global contexts. Trained an‎d validated on the CAMUS dataset, it achieves Dice Similarity Coefficients of 91.11±0.19, 87.60±0.11, an‎d 87.85±0.29 for left ventricle endocardium, epicardium, an‎d left atrium in two-chamber views, an‎d 92.25±0.17, 87.29±0.17, an‎d 91.71±0.10 in four-chamber views. These results demonstrate that our proposed model outperforms pure CNN-based models (such as UNet), pure transformer-based models (such as Swin-Unet), an‎d existing CNN-Transformer hybrid (such as UNETR-2d) architectures in the state of the art, achieving superior segmentation accuracy an‎d robustness.

يادگيري عميق , ناحيه بندي خودكار , اكوكارديوگرافي دوبعدي , شبكه عصبي كانولوشني , ترنسفورمر , معماري هيبريد

Deep Learning , Automatic Segmentation , 2D Echocardiography , Convolutional Neural Network , Transformer , Hybrid Architecture

Saeed Chamani

Hamid Behnam