Probability-Based Two-Stage CNN for Pulmonary Embolism Detection in Computed Tomography Pulmonary Angiography

Abstract

Pulmonary embolism (PE) can impede the normal flow of blood, which can result in reduced oxygen levels in vital organs and be life-threatening. Therefore, early detection and treatment of PE can effectively reduce mortality. Computed tomography pulmonary angiography (CTPA) is the clearest and most immediate tool available for PE diagnosis. With an increase in the CTPA image quality and number of slices, physician fatigue associated with interpreting these images during the diagnostic process has also increased. Therefore, this study proposes a modified two-stage convolutional neural network-based approach to automatically detect areas suspected of containing PE in CTPA images to reduce the burden of diagnosis on doctors. The main functions for eliminating false positives include principal component analysis, enlarged feature map (EFM), and probability-based anchor point extraction (PAE). Experimental results showed that although EFM can slightly improve the accuracy for small-object detection, it significantly increases the time required for training. PAE can reduce the training time and slightly improve detection accuracy. But the simultaneous use of EFM and PAE can significantly increase sensitivity (10.57%) with an increase of only 29 min in the training time.