introduction

Hypertrophic Cardiomyopathy(HCM) and Hypertensive Heart Disease(HHD) are two different diseases with similar morphological features. Differentiating HCM from HHD is of significant clinical value, yet an accurate differentiation bears great challenges, especially for patients with both HCM indications and history of hypertension. In this study, we sought to develop a deep neural network to differentiate HCM from HHD based on cardiac magnetic resonance(CMR) cine imaging. In particular, we hypothesize that dynamic information contained in cine imaging is important for accurate classification of the two diseases by neural networks.

Methods

A total of 342 HCM and HHD patients who underwent CMR exam was included in this study based on diagnosis. CMR images were acquired by experienced radiologists using a 3T MRI scanner (Ingenia, Philips) equipped with a 16-channel spine coil and 16-channel torso coil. bSSFP cine imaging was performed in short-axis view with breath-holding. Typical cine parameters were: TR/TE/flip angle/slice thickness/bandwidth/resolution/number of phases/SENSE factor = 2.7ms/1.4ms/45°/5-10mm/1900(Hz/pixel)/1.6-2mm/30/2. All images were then cropped and interpolated to 100Í100 voxels with 1.5mm isotropic resolution. Seven central slices continuously selected from the stack of cines and 16 continuous phases (Phase 1 to 16 covering systole and early diastole) were used in classification. Nine deep neural network models were trained and tested. To verify the hypothesis about cruciality of dynamic information, we divided the 9 models into static models and dynamic models, where the former generated predictions for each end-diastolic 2D image and the latter for each movie. For static models, we included pre-trained VGG16 and Resnet50 networks, which were fine-tuned by freezing the first 4 blocks and training the remaining blocks^1,2, completely-retrained VGG16 and Resnet50, and a 4-layer Vanilla CNN for comparison. For the dynamic models, we included a Resnet-like model, a 16-layer Vanilla CNN (CNN-A, Figure 1), and two simplified CNNs (CNN-B and CNN-C) respectively modified from CNN-A by replacing 8 convolutional layers with 1 pooling layer and 12 convolutional layers with 2 pooling layers. Training was performed with data augmentation including rotation (±15°) and translation(±10%), 10-fold cross validation, and with the Adam algorithm (learning rate=10-5), and batch sizes of 32-128. Each model generated a probability score for each of the 7 slices in each patient. HCM was predicted for the patient if the average score over the 7 slices was greater than 0.5. In addition to relying solely on neural networks, we also leveraged the concept of ensemble learning by combining the prediction result of each neural network model with left ventricle ejection fraction (LVEF) to generate a prediction for each patient. Three binary classification models, including logistic regression(LR), support vector machine(SVM), and K-nearest neighbor (KNN) were compared.³

Results

After excluding patients with other documented heart diseases, 303 patients including 190 HCM and 113 HHD were included for further analysis. Detailed patient characteristics information is shown in Table 1. Note that although the HCM group had a significantly lower hypertension prevalence, 45% of HCM subjects still had hypertension(p<0.001).Table 2 shows the classification accuracy of the 9 neural network models. The metric |Δ Accuracy| refers to the absolute value of HCM accuracy - HHD accuracy, indicating classification bias mostly caused by group size imbalance. Among all static models, surprisingly, the simplest 4-layer CNN static model reached the highest combined accuracy(81.2% ± 5.5%) and second-lowest |Δ Accuracy|(21.8% ± 12.2%).⁴ The performance of dynamic models was overall better than static models, suggesting dynamic information is important for accurate prediction. Within the dynamic models, CNN-A had the best diagnostic accuracy (86.8% ± 3.5%) and lowest |Δ Accuracy|(11.1%±9.0%).Table 3 shows the result of ensemble learning. The SVM method achieved better accuracy than the other two approaches. However, compared with the result obtained by networks alone, there was only a small change of accuracy (87.5% vs. 86.8%) accompanied by a higher standard deviation (6.0% vs. 3.5%).Figure 2 shows four patients, two correctly predicted and the other two incorrectly predicted by CNN-A. The correctly predicted patients had rather typical morphological and functional features, including asymmetric hypertrophy and apical obliteration for the HCM patient and symmetric hypertrophy for the HHD patient.⁵ The incorrectly predicted patients, however, showed a morphology quite opposite to general impression, with the HCM patient showing a symmetric hypertrophy and HHD an asymmetric hypertrophy.

Discussion and conclusions

The results showed that dynamic models performed better than the static models, suggesting that features about the left ventricular contraction and relaxation are important for accurate differentiation of the two diseases. A recent study including 53 HHD subjects and 108 HCM subjects performed by Ulf Neisius et.al showed 80% accuracy with use of radiomics.⁶ Our testing data accuracy is comparable and mildly higher than the previous study, potentially due to the use of deep neural networks and dynamic information. In conclusion, our result shows that convolutional neural networks based on MR cine dataset only are feasible and reasonably accurate for differentiation of HCM vs HHD.

Acknowledgements

No acknowledgement found.