Development and validation of explainable deep learning models for classification of atrial fibrillation subtypes using cardiac computed tomography

Figures

• 

  Figure 1. Overview of the proposed method. This figure illustrates the workflow of the proposed method. Cardiac CT images were acquired, preprocessed, and converted into binary masks of the left atrium. Data augmentation was applied, and a 3D convolutional neural network (3D-CNN) was trained to classify AF subtypes. Model performance was evaluated using standard metrics, and Grad-CAM was used to visualize regions contributing to model predictions. AUROC = area under the receiver operating characteristic curve, CNN = convolutional neural network, CT = computed tomography, Grad-CAM = gradient-weighted class activation mapping, NPV = negative predictive value.

• 

  Figure 2. Receiver operating characteristic curves. A: ROC curves for three 3D CNN models for classifying AF subtype in the internal validation cohort. B: ROC curves for selected echocardiographic and CT-derived imaging parameters in the internal validation cohort. C: ROC curves for 3D-CNN models in the external validation cohort. Among all models, 3D-DenseNet201 demonstrated the highest overall classification performance. Diagnosis of persistent AF was considered the positive class for all analyses. AF = atrial fibrillation. Other abbreviations are defined in Figure 1.

• 

  Figure 3. Regional distribution of Grad-CAM activation hotspots in patients with persistent AF. Heatmaps were generated using Grad-CAM from the 3D-DenseNet201 model and overlaid on LA segmentation masks. The left atrium was anatomically divided into 13 predefined regions. The percentages shown in the figure denote the proportion of cases exhibiting Grad-CAM activation in each predefined anatomical region. A: Internal validation cohort: The highest density was located in the anterosuperior wall (78.3%). B: External validation cohort: The anterosuperior wall again showed the highest hotspot frequency (72.8%). The posterior wall regions exhibited minimal activation in both internal and external validations, indicating a limited role in CNN-based classification of AF subtypes.

• 

  Figure 4. Frequency of Grad-CAM activation by anatomical region in correctly and misclassified cases. A: Internal validation cohort, anterior–posterior (AP) view. B: Internal validation cohort, posterior–anterior (PA) view. C: External validation cohort, AP view. D: External validation cohort, PA view. In both cohorts, correctly classified PerAF (blue bars) and misclassified PAF (orange bars) cases showed frequent activation in the anterosuperior wall, with minimal involvement of the posterior wall. In contrast, misclassified PerAF cases (green bars) exhibited no consistent activation pattern across the 13 predefined LA regions.