Enhancement and Lightweight Fusion U-Net for Segmentation of Cardiac Structures in MRI

Abstract

Magnetic resonance imaging technology and the segmentation process are essential parts in obtaining accurate diagnosis of cardiac diseases. Recently, deep learning networks have shown progress in achieving good accuracy. However, the 2D network may not produce the required accuracy, and the use of 3D networks is restricted because of the computation and storage complexities. In this paper, a lightweight fusion U-Net (LWFU-Net) is proposed to achieve the trade-off between segmentation accuracy and complexities. LWFU-Net is a new variant of U-Net and LVRV-Net. It enhances 2D information by using a propagation flow approach with multipath and multilevel fusion. LWFU-Net operates via three stages. First, the adjacent slices of target slice are fused by using traditional average method, generating context information. Subsequently, the target slice is segmented similar to U-Net network, where the context information is fused with the target slice and the decoding results via concatenation operator and then processes information by using convolutional layers. Finally, the deep supervision generates the final outcome. The proposed networks are evaluated on the public Automated Cardiac Diagnosis Challenge dataset by using fivefold cross validation. LWFU-Net outperforms single-path LWU-Net and agrees with the LVRV-Net. The best performance of LWFU-Net exhibits mean dice similarity scores of 0.939 left ventricular cavity, 0.861 right ventricular cavity, and 0.885 left ventricular myocardium over the subtest of 10% of training set. Results demonstrate that LWFU-Net has an acceptable usage of 51 MB (weight memory), 13,274,252 (trainable parameters), and 20 G flop. It achieves a speedup of 6 for the training and approximately 2 for the testing.