WE‐G‐204‐07: Automated Characterization of Perceptual Quality of Clinical Chest Radiographs: Improvements in Lung, Spine, and Hardware Detection

Abstract

Purpose:To develop and validate more robust methods for automated lung, spine, and hardware detection in AP/PA chest images. This work is part of a continuing effort to automatically characterize the perceptual image quality of clinical radiographs. [Y. Lin et al. Med. Phys. 39, 7019–7031 (2012)]Methods:Our previous implementation of lung/spine identification was applicable to only one vendor. A more generalized routine was devised based on three primary components: lung boundary detection, fuzzy c‐means (FCM) clustering, and a clinically‐derived lung pixel probability map. Boundary detection was used to constrain the lung segmentations. FCM clustering produced grayscale‐ and neighborhood‐based pixel classification probabilities which are weighted by the clinically‐derived probability maps to generate a final lung segmentation. Lung centerlines were set along the left‐right lung midpoints. Spine centerlines were estimated as a weighted average of body contour, lateral lung contour, and intensity‐based centerline estimates. Centerline estimation was tested on 900 clinical AP/PA chest radiographs which included inpatient/outpatient, upright/bedside, men/women, and adult/pediatric images from multiple imaging systems. Our previous implementation further did not account for the presence of medical hardware (pacemakers, wires, implants, staples, stents, etc.) potentially biasing image quality analysis. A hardware detection algorithm was developed using a gradient‐based thresholding method. The training and testing paradigm used a set of 48 images from which 1920 51×51 pixel2 ROIs with and 1920 ROIs without hardware were manually selected.Results:Acceptable lung centerlines were generated in 98.7% of radiographs while spine centerlines were acceptable in 99.1% of radiographs. Following threshold optimization, the hardware detection software yielded average true positive and true negative rates of 92.7% and 96.9%, respectively.Conclusion:Updated segmentation and centerline estimation methods in addition to new gradient‐based hardware detection software provide improved data integrity control and error‐checking for automated clinical chest image quality characterization across multiple radiography systems.