DeT-Net: A Two-stage Method for Accurate Automatic Heart Segmentation
Deng Yuhang, Hu Shuaicong, Feng Yuexiao, Jian Hua Wei, Cuiwei Yang · 2023
Heart segmentation based on clinical data is a critical task for presymptomatic cardiovascular disease diagnosis and the following sustainable medical care. The complexity of cardiac image segmentation lies in the fact that the cardiac substructures are closely located and they only occupy a small portion of clinical MRI images, making it challenging for segmentation networks to focus on the segmentation of heart. Furthermore, the shapes and sizes of these cardiac structures in images can vary significantly due to differences among patients, imaging equipment, and other factors. The insufficient segmentation accuracy poses challenges in the clinical application of cardiac image segmentation algorithms. In this paper, to improve the segmentation accuracy between heart substructures, we present a two-stage heart detection and segmentation model called DeT-Net (Detection-based TransUNet with an improved bottleneck). The proposed detection-based method improves segmentation accuracy and efficiency by identifying and focusing only on regions likely to contain the heart. Experiment on Automated Cardiac Diagnosis Challenge (ACDC) datasets proves that our method outperforms other networks with a Dice Similarity Coefficient (DSC) of 91.55%.