Optical coherence tomography–enabled classification of the human venoatrial junction

Figures

• 

  Fig. 1. Examples of B-scans, corresponding histology, and figures showcasing the intensity as a function of depth from the volumes in the dataset are shown. The dashed, pink boxes in the B-scans are 367 μm×800 μm. (a) This B-scan and corresponding Masson’s trichrome image are from donor 4. The venous tissue of the PV allows for deeper penetration of light. The endocardium is more scattering. (b) This figure showcases how the intensity of light decays with depth. The patches used in this analysis had the dimension of the dashed, pink boxes in the B-scans (∼15.3 mm×3.3 mm). The A-lines in these patches were averaged. No patches with multiple labels were used. In this volume, the light intensity decays more slowly in the pulmonary vein than in the left atrium. (c) This Masson’s trichrome image and B-scan are from donor 10. (d) In this sample, the light intensity in the pulmonary vein decays faster than that of the left atrium. All scale bars represent 1 mm.

• 

  Fig. 2. Algorithm flowchart for both the machine learning and deep learning approaches. Features selected based on feature engineering techniques were input into the machine learning algorithms. Depending on the implementation, deep learning approaches took only the patches or a combination of patches and selected features as input.

• 

  Fig. 3. Feature engineering strategy to select the features for input into classification algorithms. (a) Ten features were extracted, and the correlation matrix of the features is shown. Locations with “x” are indicative of a statistically insignificant correlation (P>0.05). The size and hue of the circles describe the magnitude of the correlation. (b) The feature importance of RF as a function of the mean decrease in impurity averaged across seven folds. The attenuation coefficient SD is the most important, whereas homogeneity is the least important feature. (c), (d) The ROC curves and the area under the ROC curves of the top features and the attenuation coefficient. All of these features have an area under the ROC curve over 0.6, indicating discriminating power. The final selected features were correlation, skewness, attenuation coefficient SD, and energy.

• 

  Fig. 4. Distribution of average feature values across venoatrial junctions, N=26, for the 235 μm×800 μm patches. The line represents the mean ± standard deviation. T-testing indicates that the means are significantly different. * indicates P≤0.05, ** represents P≤0.01, and *** represents P≤0.001.

• 

  Fig. 5. ROC curves across the sevenfold test set for all methods: (a) logistic regression (machine learning), (b) random forest (machine learning), (c) patch-only CNN (deep learning), (d) ensemble CNN V1 (deep learning), and (e) ensemble CNN V2. Splits 1 to 6 had 4 PV-LA junctions in the test set, whereas split 7 had only two samples of PV-LA junctions in the test set to avoid the repetition of samples. Classification performance varied depending on the implementation; however, all algorithms performed best on splits 2 and 7 and had the worst performance on split 5. AUROC, area under the receiver operating characteristic curve.

• 

  Fig. 6. En face image (a) and probability maps for the methods: (b) RF (machine learning), (c) patch-only CNN (deep learning), and (d) ensemble CNN V1 (deep learning) of a sample from donor 4 (split 2). The ensemble CNN deep learning version 1 (d) classified the PV most confidently of all the algorithms. The CNN architecture that took patches as input also performed well, but the class probabilities across the volume were not as high. Samples with prediction probabilities of 0.5 and higher are classified as PV. The color bar is the probability of PV classification. The AUROC curve for CNN ensemble version 1 for this volume is 97.5%. The scale bars are 1 mm.

• 

  Fig. 7. En face image (a) and probability maps for the methods: (b) RF (machine learning), (c) patch-only CNN (deep learning), and (d) ensemble CNN V1 (deep learning) from donor 4 (split 6). The CNN with patches only (c) and the ensemble CNN (d) classified the PV most confidently. The patches provide the CNNs additional contrast that random forest does not have, leading to fewer positive predictions. In these maps, the patch CNN (c) classifies fewer LA patches as PV than the ensemble CNN that takes patches as input (d). The color bar is the probability of PV classification. Samples with prediction probabilities of 0.5 and higher are classified as PV. The AUROC curve for CNN ensemble version 1 for this volume is 74.8%. The scale bars are 1 mm.

• 

  Fig. 8. Sample correct classifications of OCT image patches from a region within the PV [(a), (d)] and a region within the LA [(g), (k)], and corresponding CNN GradCAM visualizations for the LA [(b), (e), (g), (k)] and PV [(c), (f), (i), (l)] gradients. The importance of pixels in the patch for class prediction is highlighted within these GradCAM maps. Higher-intensity pixels indicate regions with higher gradients, which are more impactful to classification than lower-intensity regions. For classifying the PV samples [(a), (d)], GradCAM shows that the PV gradient [(c), (e)] is sensitive to the top of the patch and notices the gradual intensity falloff. For classifying LA [(g), (j)], GradCAM shows that the LA gradient [(h), (k)] identifies the abrupt falloff of intensity at the endocardium-myocardium transition.