CMMCSegNet: Cross-Modality Multicascade Indirect LGE Segmentation on Multimodal Cardiac MR

Figure 6

Qualitative comparisons of our CMMCSegNet with the other four state-of-the-art CNN-based segmentation methods, where the LSGAN adversarial loss and vgg perceptual loss are employed and the loss weight parameters is manually given. (a) Direct segmentation, from left to right: LGE, ground truths with zoom-in views and prediction results with zoom-in views using FCNs, U-net, U-net++, and Attention U-net for segmentation on real LGE modality; (b) indirect segmentation, from left to right: LGE, fake bSSFP, ground truths with zoom-in views, and prediction results with zoom-in views using FCNs, U-net, U-net++, Attention U-net, and our CMMCSegNet for segmentation on fake bSSFP modality translated from LGE modality.

(a) Direct segmentation for the LGE modality

(b) Indirect segmentation for fake bSSFP translated from LGE modality