A deep-learning approach for myocardial fibrosis detection in early contrast-enhanced cardiac CT images

Figures

• 

  Figure 1. Workflow of the proposed approach for myocardial fibrosis detection. For training and testing purposes, for each early CE-CCT images the myocardial region was extracted and divided in sectors, according to the AHA model. Then scar ground-truths obtained with manual tracing of myocardial fibrosis contours on late CCT images were used to label each corresponding sector on early phase. Abbreviations as in Figure 1. AHA, American Heart Association; CCT, cardiac computed tomography; CMR, cardiac magnetic resonance; CE, contrast-enhanced.

• 

  Figure 2. Example of hypoenhanced area on early and late contrast-enhanced CCT and corresponding contrast-enhanced CMR image. Abbreviations as in Figure 1.

• 

  Figure 3. Receiver operating characteristic (ROC) curve for each fold of the 5-fold cross-validation, and the mean. ROC, Receiver Operating Characteristic; AUC, Area Under the Curve.

• 

  Figure 4. Examples of model’s classification for a basal, mid-ventricular and apical slice. From left to right: ground truth, raw image, scar segmentation and the predicted classification. CNN, convolutional neural network; other abbreviations as Figure 1.

• 

  Figure 5. Per-segment analysis of the concordance (Cohen’s K statistic) between CE-CMR and CNN detection on early CE-CCT. CI: confidence interval.

• 

  Figure 6. 16-segment AHA model comparison between CE-CMR visual analysis and CNN detection on early CE-CCT for patients with ischemic and non-ischemic LV dysfunction: red and green represent pathological and healthy sectors, respectively. Blu indicates the sector excluded during pre-processing phase. Abbreviations as in Figure 1.