Atrial scar quantification via multi-scale CNN in the graph-cuts framework

Figures

• 

  Fig. 1. The challenges of automatic delineation of scars from LGE MRI: (a) two typical LGE MRIs with poor quality; (b) thin atrial walls highlighted using bright white color in the figure; (c) surrounding enhanced regions pointed out by the arrows, where (1) and (2) respectively indicate the enhanced walls of descending and ascending aorta, (3) denotes the enhanced walls of right atrium.

• 

  Fig. 2. Flowchart of the proposed framework for LA scar quantification and analysis.

• 

  Fig. 3. Pipeline of the projection and patch extraction phases: (a) target LGE MRI; (b) WHS result of LGE MRI propagated from Ana-MRI; (c) extracted LA and PV from WHS; (d) LA endocardial surface after projection; (e) patch along the normal direction of the LA endocardial surface; (f) patches with random shift in the training phase.

• 

  Fig. 4. Construction of the graph and the explicit learning of the graph potentials by MS-CNN (MSPs are integrated and represented using red cuboid in this work). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

• 

  Fig. 5. The hierarchical architecture of the networks: (a) T-NET: the input is the MSPs of the target node, and the output is the predicted t-link weight; (b) N-NET: the input is the patch pair of two neighbor nodes, and the output is the predicted n-link weight; (c) Patch-NET: the input is a patch, and the output is the patch features. Note that the diagram takes 13 × 13 × 17 patch as an example.

• 

  Fig. 6. Scatter point plot for analyzing the correlation between the LA segmentation performance and scar quantification accuracy, both indicated by Dice scores. The Pearson coefficient and Spearman’s rank coefficient are respectively 0.1412 and 0.0110.

• 

  Fig. 7. Dice scores of the proposed method with different parameterizations: (a) performance against different patch sizes (λ=0.6); (b) performance against different values of the balancing parameter λ to weight the t-link and n-link terms in the graph-cuts framework; (c) performance against different random shift ranges R; (d) performance against different numbers of scales Ns.

• 

  Fig. 8. Boxplots of the Dice scores of scar segmentation by the nine methods.

• 

  Fig. 9. 3D visualization of the LA scar classification results using the nine methods. This is the median case selected from the test set in terms of Dice score of scars by the proposed method. The scarring areas are red-colored on the LA surface mesh, which can be constructed either from LAM (LA surface in white) or from LAauto (LA surface in light yellow). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

• 

  Fig. 10. Axial view of the images, the ground truth scar segmentation and the results by the proposed method. The red and white color labels represent the scar and normal wall, respectively. Arrow (1), (2) and (3) indicate the major classification errors of the proposed method caused by the surrounding enhanced regions, respectively from the right atrium wall, ascending aorta wall and descending aorta wall; arrow (4) shows an error from the misalignment between the automatic LA segmentation and the ground truth; arrow (5) illustrates that the proposed method can still perform well, even though the automatic LA segmentation contains obvious errors. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)