The effect of pulmonary vessel suppression on computerized detection of nodules in chest CT scans

Abstract

In chest computed tomography (CT) scans, pulmonary vessel suppression can make pulmonary nodules more evident, and therefore may increase the detectability of early lung cancer. The purpose of this study was to develop a computer-aided detection (CAD) system with a vessel suppression function and to verify the effectiveness of the vessel suppression on the performance of the pulmonary nodule CAD system. Methods: A CAD system with a vessel suppression function capable of suppressing vessels and detecting nodules was developed. First, a convolutional neural network (CNN)-based pulmonary vessel suppression technique was employed to remove the vessels from lungs while preserving the nodules. Then, a CNN-based pulmonary nodule detector was utilized to sequentially generate nodule candidates and reduce false positives (FPs). The performance levels of CAD systems with and without the vessel suppression function were compared using 888 three-dimensional chest CT scans from the Lung Nodule Analysis 2016 (LUNA16) dataset. The pulmonary nodule detection results were quantitatively evaluated using the average sensitivity at seven predefined FP rates: 0.125, 0.25, 0.5, 1, 2, 4, and 8 FPs per scan. Results: The developed pulmonary nodule CAD system improved the average sensitivity to 0.977 from 0.950 owing to the addition of the vessel suppression function. Conclusions: The vessel suppression function considerably improved the performance of the CAD system for pulmonary nodule detection. In practice, it would be embedded in CAD systems to assist radiologists in detecting pulmonary nodules in chest CT scans.

PUNDIT: Pulmonary nodule detection with image category transformation

“…Their method achieved an average sensitivity of 0.891 at seven FP rates in LUNA16 dataset. Gu et al 9 used a pulmonary nodule detector based on two feature pyramid networks and achieved an average sensitivity of 0.950 at seven FP rates in LUNA16 dataset. Luo et al 10 proposed 3D sphere representation-based center-points matching network (SCPM-Net), which was an anchor-free detection network using sphere representation and center points matching.…”

“…Their method achieved an average sensitivity of 0.891 at seven FP rates in LUNA16 dataset. Gu et al 9 . used a pulmonary nodule detector based on two feature pyramid networks and achieved an average sensitivity of 0.950 at seven FP rates in LUNA16 dataset.…”

Lung Cancer Screening and Nodule Detection: The Role of Artificial Intelligence

“…If vessels are removed from the lung while preserving nodules completely, nodule detection could be much easier for both human readers and AI, with less falsepositive results. Inspired by this idea, Gu et al developed a deep learning-based detection system using the pulmonary vessel suppression technique, and explored the effect of removal of vessels on the performance of AI [47]. The ablation experiment demonstrated that sensitivity improved from 96.9% to 98.6%, with a false-positive rate that dropped from 7.65 to 0.92 per scan after ruling out pulmonary vessels for the same candidate detection and false-positive reduction algorithms.…”