Segmentation of the Left Ventricle on EchoCG Images Using MultiResUnet
Andrey Sheka, Victor Samun, Tatiana V. Chumarnaya, Olga E. Solovyova · 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON) · 2019
We compared the segmentation quality obtained with the following network architectures: Unet, Wide Unet, Unet++ and MultiResUnet. The MultiResUnet architecture has shown improved quality in other medical image segmentation tasks. It is applied to the problem of left ventricular segmentation for the first time. This architecture showed segmentation accuracy on augmented data equal to 92.78% by Dice metric. This is 1.3% more than the result of Unet++ and 2.5% more than the result of Unet. Also, this architecture showed a smaller variance of segmentation accuracy on cross validation: 1.2% vs. 1.4%.