Multi-CBCS: Multimodal Contrast Bolus Consistency Network for Pulmonary Embolism Detection by Integrating CTPA and Lung Perfusion Imaging

Abstract

Pulmonary embolism (PE) is a dangerous heart disorder that should be effectively diagnosed as soon as possible. The clinical standard of detecting emboli based on contrast filling defects is computed tomography pulmonary angiography (CTPA), and lung perfusion scintigraphy (V/Q scan) is a complementary test of the pulmonary blood flow. Nevertheless, the current deep learning methods are mainly based on one-modality image analysis, and they do not explicitly simulate the dynamics of contrast bolus or combine the functional perfusion information, which restricts the diagnostic strength and clinical readability. Furthermore, the Contrast Bolus Consistency Module (CBCM) increases interpretability by accounting for spatial consistency of contrast flow through pulmonary arteries, allowing the network to emphasize meaningful interruptions of this process caused by embolic blockages, which regular single modality CNNs cannot due to implicit feature extraction. Addressing these drawbacks, a Contrast Bolus Consistency Network is suggested to detect multimodal pulmonary embolism based on the images of CTPA and lung perfusion scintigraphy. It is based on a convolutional neural network (CNN)-based backbone that has been developed on the basis of which hierarchical feature extraction is enabled. The presented framework is able to integrate physiologically coherent multimodal features by uniting abnormalities of anatomy contrast with functional deficits of perfusion by using cross-modal attentional processes. A specific contrast bolus modeling module evaluates consistency of spatial attenuation along pulmonary arteries to detect contrast discontinuity that is a sign of embolic obstruction. The results of the experiment prove that the combination of multiple modalities CBC-Net is much better, 99.1% accuracy, 98.4% precision, 98.4% recall and 95.7% F1-score according to the experimental outcomes are significantly higher when compared to existing single-modal methods. These results indicate that there is clinical potential in combining structural and functional imaging modalities as a viable method of detecting pulmonary embolism.