A Two-Stage Convolutional Neural Network for Pulmonary Embolism Detection From CTPA Images

Xin Yang, Yi Lin, Jianchao Su, Xiang Wang, Xiang Li, Jingen Lin, Kwang-Ting ChengIEEE Access 2019 · DOI

Abstract

Pulmonary Embolism (PE) is a critical medical condition where blood clots obstruct lung arteries, requiring rapid and precise diagnosis. This paper introduces a high-performance, two-stage deep learning framework designed to detect PE in Computed Tomography Pulmonary Angiography (CTPA) scans. The system combines a 3D proposal network with a novel vessel-alignment technique for 2D classification. Experimental results demonstrate a 75.4% sensitivity rate at 2 false positives per scan, significantly outperforming previous state-of-the-art models and offering a robust tool for clinical decision support.

Introduction

Pulmonary Embolism (PE) occurs when a blood clot lodges in the lung's arteries, a condition that can be rapidly fatal if missed. To diagnose it, clinicians utilize Computed Tomography Pulmonary Angiography (CTPA), which generates massive 3D volumes consisting of 300 to 500 slices. For radiologists, identifying a PE is a high-stakes task akin to finding a needle in a haystack; they must trace every branching artery to spot tiny "filling defects"—dark spots where clots block the contrast agent. This process is exhausting and prone to human error. While automated Computer-Aided Detection (CAD) systems exist, they historically suffer from high false-positive rates, often becoming a clinical burden. This research aims to bridge this gap by creating a detector that is both highly sensitive and precise.

Methodology

The proposed framework employs a "divide and conquer" strategy using a two-stage Convolutional Neural Network (CNN) architecture.

Stage 1: 3D Candidate Proposal Subnet

Instead of analyzing every voxel, the system uses a 3D Fully Convolutional Network (FCN) to identify potential regions of interest. Inspired by the Faster R-CNN architecture, it utilizes "anchor cubes" of various sizes to suggest potential PE locations. A key innovation is the incorporation of a 3D location map; because PEs are more likely to occur at specific anatomical landmarks like arterial bifurcations, providing the network with (x, y, z) coordinates allows it to learn high-risk zones.

The two-stage CNN architecture: Stage 1 proposes 3D candidates using anatomical location information, while Stage 2 performs false positive removal via aligned 2D classification.

Stage 2: False Positive Removal via Vessel Alignment

The "secret sauce" of this approach is a novel vessel-alignment technique. Clots appear differently depending on the vessel's orientation, which often confuses standard classifiers. The authors use Principal Component Analysis (PCA) to determine the orientation of the affected blood vessel, then rotate and align the 3D candidate cube so the vessel runs straight through it. Finally, three aligned 2D cross-sections (Coronal, Traverse, and Sagittal) are extracted and fed into a ResNet-18 classifier to confirm the presence of a clot.

The vessel alignment pipeline: candidates are cropped, binarized, and oriented via PCA to produce standardized 2D slices for the final classifier.

Results and Key Insights

The study utilized two primary datasets: the public CAD-PE Challenge (40 scans) and the PE129 Dataset (129 scans with 269 embolisms). The two-stage approach showed dramatic improvements over existing methods:

  • Superior Sensitivity: On the public challenge dataset, the model achieved 75.4% sensitivity at 2 false positives per scan, compared to the previous state-of-the-art of 60.8%.
  • Robustness: On the internal PE129 dataset, the model reached up to 84.2% sensitivity at a 5mm error tolerance.
  • Impact of Alignment: Ablation studies proved that the vessel alignment step alone improved sensitivity by 20-30% compared to using standard, unaligned views.

Discussion and Conclusion

The research highlights that while 3D context is vital for finding potential clots, properly aligned 2D cross-sections are highly efficient for verification. By mathematically aligning the "camera view" to the vessel's longitudinal axis, the model focuses on the actual features of the clot rather than random anatomical orientation.

Despite these gains, challenges remain in detecting very small "subsegmental" embolisms in peripheral arteries. Future work will focus on cross-center validation with diverse scanner protocols and optimizing the pipeline for real-time integration into radiologist workflows.

Summary updated

BibTeX
@article{yang2019two,
  author = {Xin Yang and Yi Lin and Jianchao Su and Xiang Wang and Xiang Li and Jingen Lin and Kwang-Ting Cheng},
  title = {{A Two-Stage Convolutional Neural Network for Pulmonary Embolism Detection From CTPA Images}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2925210},
}
Figure