Year-end Sale % OFF 
Abstract
This paper presents a two-stage convolutional neural network (CNN) for automated detection of pulmonary embolisms (PEs) on CT pulmonary angiography (CTPA) images. The first stage utilizes a novel 3D candidate proposal network that detects a set of cubes containing suspected PEs from the entire 3D CTPA volume. In the second stage, each candidate cube is transformed to be aligned to the direction of the affected vessel and the cross-sections of the vessel-aligned cubes are input to a 2D classification network for false positive elimination. We have evaluated our approach using both the test dataset from the PE challenge and our own dataset consisting of 129 CTPA data with a total of 269 embolisms. The experimental results demonstrate that our method achieves a sensitivity of 75.4% at two false positives per scan at 0 mm localization error, which is superior to the winning system in the literature (i.e., sensitivity of 60.8% at the same level of false positives and localization error). On our own dataset, our method achieves sensitivities of 76.3%, 78.9%, and 84.2% at two false positives per scan at 0, 2, and 5 mm localization error, respectively.
Highlights
Pulmonary Embolism (PE) refers to the situation when a blood clot becomes lodged in one of the arteries that go from the heart to the lungs
Existing methods typically consist of two steps: 1) detecting a list of candidates from an entire computed tomography pulmonary angiography (CTPA) volume based on voxel-level features, and 2) removing false positives from candidates based on region-level features and a classifier [3], [4]
We have evaluated our approach using the entire 20 CTPA test dataset from the PE challenge [9], achieving a sensitivity of 75.4% at 2 false positives per scan at 0mm localization error
Summary
Pulmonary Embolism (PE) refers to the situation when a blood clot becomes lodged in one of the arteries that go from the heart to the lungs. Existing methods typically consist of two steps: 1) detecting a list of candidates from an entire CTPA volume based on voxel-level features, and 2) removing false positives from candidates based on region-level features and a classifier [3], [4]. Extracted handcrafted features based on CT values, local contrast and the second derivatives of voxels for candidate detection and leveraged the volume, effective length and mean local contrast of grouped voxels as region-level features for false positive removal. Due to the limited representation ability of these handcrafted features, conventional methods often suffer from a high false positive rate in order to achieve an acceptable sensitivity. Our PE detection network is a cascade of two stages: 1) a 3D candidate proposal subnet based on a 3D fully convolutional neural network (FCN), and 2) a false positive removal subnet based on vessel-aligned candidate transformation and a 2D classification network. The first stage is a 3D fully convolutional network that proposes candidate, and the second stage extracts vessel-aligned 3D candidate cubes and removes false positives based on 2D cross-sections of vessel-aligned cubes and a ResNet-18 classifier [10]
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
