Deep Learning-Based Classification of Pulmonary Hypertension Using Early-Fused CTPA and Pulmonary Artery Segmentation Maps.
文献摘要
Computed tomography pulmonary angiography (CTPA) is vital for diagnosing pulmonary hypertension (PH), especially thromboembolic pulmonary hypertension (CTEPH). However, manual image interpretation of CTPA often leads to misdiagnosis. This study aims to automatically segment the pulmonary artery from CTPA images and develop a deep learning model by fusing CTPA images with pulmonary artery segmentation results to distinguish pulmonary hypertension (PH vs. subjects without PH (no-PH)) and its PH subtype, CTEPH (CTEPH, other PH vs. no-PH). The study included 92 patients with CTEPH, 85 patients with other types of PH, and 140 no-PH. We completed the image preprocessing and segmentation of the pulmonary arterial tree of all the sample data. Based on four deep learning model classifiers-ResNet50, DenseNet, SENet, and Swin Transformer, we respectively developed four distinct network architectures: Single input channel for CTPA images, as well as early, middle, and late fusion strategies that combine CTPA images with their corresponding pulmonary vascular segmentation maps. The performance of the testing experiments has shown that the DenseNet (CTPA-seg) early fusion model achieves optimal performance in the binary classification task of pulmonary hypertension (PH vs. no-PH), with an area under the curve (AUC) of 0.927. Meanwhile, the Swin Transformer (CTPA-seg) early fusion model shows the best performance in the ternary classification of pulmonary hypertension subtypes (CTEPH, other PH vs. no-PH), achieving an AUC of 0.855. Compared with models that only use clinical images, the early feature fusion strategy shows significantly higher accuracy in classifying and diagnosing PH, which is beneficial for the differential diagnosis of CTEPH.