Hemodynamic classification of pulmonary hypertension using CT images with machine deep learning

Pulmonary hypertension (PH) is a progressive condition characterized by high blood pressure in the pulmonary vasculature, leading to slow blood flow and increased pressure due to vascular narrowing and stiffening. Early detection of PH is crucial to prevent complications such as heart failure. Therefore, this study proposes a novel automatic classification model based on deep feature extraction to achieve an early and accurate diagnosis of PH. In this study, we collected computed tomography (CT) images for PH detection and this dataset contains 807 CT images. We developed a new automatic classification architecture, called EfDenseNet, for PH detection and tested it using the collected data. We utilized pre-trained deep network architectures, namely EfficientNetb0 and DenseNet201, as feature extractors. To enhance the performance of the feature extractors, we segmented the CT images into patches and extracted deep features from both the patches and the original image. The generated features from each patch have been merged. The presented EfDenseNet generates four feature vectors by deploying the feature extraction layers (fully connected layer and global average pooling layers) of the EfficientNetb0 and DenseNet201. We used neighborhood component analysis (NCA), ReliefF, and Chi-Square (Chi2) to select the most informative features from the feature vectors, resulting in 45 different selected feature vectors. We employed support vector machine (SVM) and k-nearest neighbor (kNN) algorithms to classify the features in the classification phase. Finally, we applied a mode function-based iterative majority voting to the classification results to obtain generalized classification results. The proposed self-organized EfDenseNet architecture produced 90 classification results and 178 voted results. We selected the best result from the 178 votes and achieved the highest classification accuracy for all four classes. In addition, we validated all results using k-fold cross-validation and obtained an accuracy of 97.27% on the new PH dataset we collected, demonstrating our proposed method's high classification performance. Our proposed automatic classification model based on deep feature extraction achieved accurate and early diagnosis of PH with high classification performance. The results of this study suggest that the proposed self-organized EfDenseNet architecture is a promising approach for PH classification. The results of this study may provide additional risk stratification indicators in the future with non-invasive CTPA data. Additionally, our study gives us ideas about which patient to refer to for right heart catheterization. Keywords: Artificial intelligence; pulmonary hypertension; computed tomography; mean pulmonary arterial pressure