Segmentation Improvement of Cardiac Regions in MRI using Hybrid Early-Late Fusion U-Net (HELFU-Net)

Image fusion and the U-Net architecture have been successfully applied to many real applications, in particular, cardiac segmentation. This paper suggests a new version named Hybrid Early-Late Fusion U-Net(HELFU-Net) to segment the cardiac structure into regions. The design has been built by extending the U-Net with five encoder branches and one decoder branch. The encoder branches take advantage of the adjacency property within the cardiac slice-images stack to boost the accuracy of the target image. The first branch of the encoder in the U-Net is to merge adjacent images using the concept of early-stage fusion. The next three branches apply late-stage fusion to the features of adjacent slices that are processed separately. The last branch is for the target slice of the image, while the decoder branch retrieves the data. This design boosts spatial information collection using only 2D image slices. HELFU-Net is evaluated using a public dataset of the ACDC challenge. The experimental results gave mean dice coefficients of 0.942, 0.856, and 0.893 for left ventricular cavity, right ventricular cavity, and left ventricular myocardium, respectively, on the test dataset. Additionally, the suggested HELFU-Net gives 94.9% comparable predicted accuracy on the test dataset over a test time of 1.065 sec.