Harnessing TernausNet For The Segmentation Of Heart From Chest X-ray For The Recognition Of Atrial Septal Defect

Semantic segmentation of chest X-rays can aid in the early detection of cardiac conditions, including atrial septal defect, which improves patient outcomes. The U-Net architecture is a popular approach to semantic segmentation, but it has some limitations such as the risk of vanishing gradients and the difficulty of capturing global context information. The TernausNet architecture is an improvement over U-Net that addresses these limitations by using residual connections, VGG16 encoders, stretching convolutions, and batch normalization. In this paper, we compare the performance of TernausNet and U-Net on the cardiomegaly datasets showing that TernausNet achieves better segmentation accuracy. We also analyze the advantages of TernausNet over U-Net in terms of architectural design and explain how these advantages improve the performance of the model.