Computer-aided detection of endobronchial valves using volumetric CT

Abstract

Rationale and objectives: The ability to automatically detect and monitor implanted devices may serve an important role in patient care by aiding the evaluation of device and treatment efficacy. The purpose of this research was to develop a system for the automated detection of one-way endobronchial valves that were implanted for less invasive lung volume reduction.

Materials and methods: Volumetric thin-section computed tomographic data was obtained for 194 subjects; 95 subjects implanted with 246 devices were used for system development and 99 subjects implanted with 354 devices were reserved for testing. The detection process consisted of preprocessing, pattern recognition based detection, and a final device selection. Following the preprocessing, a set of classifiers was trained using AdaBoost to discriminate true devices from false positives. The classifiers in the cascade used two simple features (either the mean or maximum attenuation) of a local region computed at multiple fixed landmarks relative to a template model of the valve.

Results: Free-response receiver-operating characteristic analysis was performed for the evaluation; the system could be set so the mean sensitivity was 96.5% with a mean of 0.18 false positives per subject. If knowledge of the number of implanted devices were incorporated, the sensitivity would be 96.9% with a mean of 0.061 false positives per subject; this corresponds to a total of 12 false negatives and six false positives for the 99 subjects in the test dataset.

Conclusion: Software was developed for automated detection of endobronchial valves on volumetric computed tomography. The proposed device modeling and detection techniques may be applicable to other devices as well as useful for evaluation of treatment response.

Figure 1

CT image of a valve that migrated into the posterior parenchyma (top, bone window/level). An image of another valve placed too distal for complete occlusion of all segmental airways of a lobe in also shown (bottom, lung window/level).

Figure 2

Illustration of the device model, control points, and control groups. A 2D device model and examples are shown for clarity, but a 3D device model with 58 control points rotated at 98 orientations is applied on the volumetric CT data.

Figure 3

Illustration of the underlying control points of the weak classifiers selected by the algorithm for the first level of the cascade classifier. Three of the points were on the outer ring of the device model while one was inside the device lumen.

Figure 4

Algorithm for determining the confidence score for a device, which is based on the maximum attenuation feature at each control point and whether or not the control point was model as being part of a control group on the edge of the device or inside the device lumen. The confidence score is formed by the summation of all the scores for all control points.

Figure 5

Implementation algorithm for processing through the test points and generating the final set of devices.

Figure 6

Illustration of the detection process. The four levels of the cascade are used to reject non-device point templates; the majority of instances are rejected after the first cascade. More difficult instances are rejected in later cascades. The final set of devices is formed by accepting devices from the list in descending order according to their confidence score; overlaps are rejected.

Figure 7

FROC curves for the device detection performance. Curve (a) is when the system did not know how many devices were implanted; curve (b) is when the system was told how many devices should be detected.

Figure 8

Example of two correctly detected devices in collapsed lung. Note the difficulty in differentiating collapsed lung from nearby tissue. The slice thickness/interval for this series was 0.6/0.5 mm.

Figure 9

Example of two missed device anterior to two found devices. The slice thickness/interval (1.25/0.7 mm) and orientation of the devices may have resulted in partial volume averaging, resulting in the devices being missed by the initial threshold.

Figure 10

False positives were from bone, calcium deposits, and metal artifacts. The pattern of the streak artifact likely resulted in the false positive. The slice thickness/interval for this series was 1.5/0.8 mm.