DSBAV-Net: Depthwise Separable Bottleneck Attention V-Shaped Network with Hybrid Convolution for Left Atrium Segmentation

Accurate and precise segmentation of the left atrium (LA) is crucial in the early diagnosis and treatment of atrial fibrillation (AF), which is the most common heart rhythm disease in cases. The size of fibrotic tissue in patients with AF is based on manual examination of images obtained from the gadolinium-enhanced cardiac magnetic resonance imaging (MRI) technique. However, manual examination of the acquired images is time-consuming and has many difficulties, such as LA thickness between observers and resolution according to MR devices. To overcome the challenges of manual segmentation of images obtained from MRI devices, end-to-end, fully automated deep learning-based segmentation architectures have become extremely important today. In this study, an encoder–decoder-based V-shaped deep learning architecture is proposed for precise segmentation of LA. In the proposed architecture, standard convolution and depthwise separable convolution are used together. Thus, sparsely connected blocks with fewer parameters and deeply separable convolutions learn the feature representations better, increasing the robustness of the model. In addition, the bottleneck attention module has been added to each encoder layer, allowing the network to learn which features to focus on and which features to suppress in images by attention mapping channel and spatially. The proposed architecture obtained 0.915 dice and 0.844 Jaccard scores in the STACOM 2018 challenge dataset. The obtained results draw attention to the robustness of the model.