Cardiopulmonary Imaging Original Research

Transcription

1 Cardiopulmonary Imaging Original Research Ebner et al. MIP and CAD Algorithms in CT for Lung Cancer Screening Cardiopulmonary Imaging Original Research Lukas Ebner 1 Justus E. Roos 1 Jared D. Christensen 1 Tomas Dobrocky 2 Lars Leidolt 2 Barbara Brela 2 Verena C. Obmann 2 Sonya Joy 2 Adrian Huber 3 Andreas Christe 2 Ebner L, Roos JE, Christensen JD, et al. Keywords: computer-aided detection, CT, maximum intensity projection, radiation dose DOI: /AJR Received September 17, 2015; accepted after revision February 22, L. Ebner was supported by the Swiss National Science Foundation (Early Post-Doc Mobility scholarship grant). Based on a presentation at the 2015 annual meeting of the Society of Thoracic Radiology, Carlsbad, CA. 1 Department of Radiology, Duke University Medical Center, Box 3808, Durham, NC Address correspondence to L. Ebner (lukas.ebner@duke.edu). 2 Department of Diagnostic, Interventional and Pediatric Radiology, University Hospital Inselspital, Bern, Switzerland. 3 Department of Cardiovascular and Interventional Radiology, University Hospital Pitié-Salpêtrière, Paris, France. This article is available for credit. AJR 2016; 207: X/16/ American Roentgen Ray Society Maximum-Intensity-Projection and Computer-Aided-Detection Algorithms as Stand-Alone Reader Devices in Lung Cancer Screening Using Different Dose Levels and Reconstruction Kernels OBJECTIVE. The objective of our study was to evaluate lung nodule detection rates on standard and microdose chest CT with two different computer-aided detection systems ( SyngoCT-CAD, VA 20, Siemens Healthcare [CAD1]; Lung CAD, IntelliSpace Portal DX Server, Philips Healthcare [CAD2]) as well as maximum-intensity-projection (MIP) images. We also assessed the impact of different reconstruction kernels. MATERIALS AND METHODS. Standard and microdose CT using three reconstruction kernels (i30, i50, i70) was performed with an anthropomorphic chest phantom. We placed 133 ground-glass and 133 solid nodules (diameters of 5 mm, 8 mm, 10 mm, and 12 mm) in 55 phantoms. Four blinded readers evaluated the MIP images; one recorded the results of CAD1 and CAD2. Sensitivities for CAD and MIP nodule detection on standard dose and microdose CT were calculated for each reconstruction kernel. RESULTS. Dose for microdose CT was significantly less than that for standard-dose CT ( msv vs 1.65 msv; p < ). CAD1 delivered superior results compared with CAD2 for standard-dose and microdose CT (p < ). At microdose level, the best stand-alone sensitivity (97.6%) was comparable with CAD1 sensitivity (96.0%; p = 0.36; both with i30 reconstruction kernel). Pooled sensitivities for all nodules, doses, and reconstruction kernels on CAD1 ranged from 88.9% to 97.3% versus 49.6% to 73.9% for CAD2. The best sensitivity was achieved with standard-dose CT, i50 kernel, and CAD1 (97.3%) versus 96% with microdose CT, i30 or i50 kernel, and CAD1. MIP images and CAD1 had similar performance at both dose levels (p = and p = 0.48). CONCLUSION. Submillisievert CT is feasible for detecting solid and ground-glass nodules that require soft-tissue kernels for MIP and CAD systems to achieve acceptable sensitivities. MIP reconstructions remain a valuable adjunct to the interpretation of chest CT for increasing sensitivity and have the advantage of significantly lower false-positive rates. L ung cancer screening with lowdose CT has shown favorable patient outcomes in national and international trials with strong evidence to support expansion of lung CT screening into clinical practice for appropriate patients at high risk for developing lung cancer [1 5]. As lung CT screening becomes more widely available, optimizing techniques for low-dose imaging protocols [6 8] is critical to minimize radiation exposure in the screening population while preserving diagnostic accuracy for both solid and subsolid nodules. Current imaging guidelines for low-dose CT screening by the American College of Radiology (ACR) recommend a volume CT dose index of less than 3.0 mgy for a standard-sized patient [9]. Recent studies have showed that further radiation dose reduction is possible. According to Mettler et al. [10], the mean effective dose for posteroanterior and lateral radiographs of the chest is 0.1 msv (range, msv) [10]. The radiation dose of microdose chest CT examinations varies from 0.06 to 0.16 [6, 8]. At this dose level, chest CT can depict pulmonary nodules with an effective radiation dose comparable with that of a twoview chest radiograph [11]. Microdose technique introduces inherent image degradation (increased noise levels); however, these limitations can be mitigated by different processing techniques to facilitate noise reduction and nodule detection. One approach to improving image quality for microdose CT is the application of it- 282 AJR:207, August 2016

2 MIP and CAD Algorithms in CT for Lung Cancer Screening erative reconstruction models during image processing. Reconstruction kernels for microdose CT should differ from conventional kernels used in standard-dose CT because of differences in noise and spatial resolution between techniques. New data suggest that lower, softer reconstruction kernels may be favored for microdose CT in comparison with sharp lung kernels commonly used in routine low dose or diagnostic chest CT [6]. Maximum-intensity-projection (MIP) reconstruction is a relatively simple postprocessing method applied to chest CT data that has proven to increase nodule detection rates and temporal detection efficiency in both standard- and low-dose chest CT [12]. In addition, many vendors offer software solutions for assisting the radiologist with CT screening for lung cancer by automated lesion detection. Computer-aided detection (CAD) systems rely on different calculating algorithms with varying accuracy in lesion detection rates [13, 14]. Multiple studies have investigated the utility of CAD in CT screening for lung cancer and have found that CAD software improves diagnostic sensitivity and specificity of screening nodule detection [14, 15]. Several studies have detailed the effects of radiation dose, reconstruction kernel, MIP, and CAD on lung CT image quality and nodule detection [16 18]. However, to our knowledge, no comprehensive investigation has included all of these multivariate factors and examined their effect on diagnostic accuracy. The purpose of our study was to evaluate the lung nodule detection rate on standard-dose and microdose chest CT using two different CAD systems (SyngoCT-CAD, VA 20, Siemens Healthcare [CAD1]; Lung CAD, IntelliSpace Portal DX Server, Philips Healthcare [CAD2]) and MIP-processed images as well as the effect of variable reconstruction kernels. Materials and Methods Local ethics committee approval was waived because data were collected solely from anthropomorphic chest phantoms (chest phantom N1, Kyoto Kagaku). Figure 1 shows the basic study design. The phantom represents an anatomically accurate life-sized model of an male human torso. The phantom features a synthetic heart, trachea, mediastinum, and lung architecture. The soft-tissue structures and the incorporated synthetic osseous structures are comparable with human tissues, providing similar absorption rates on CT. The lungs contained embedded nodules that varied in size and density. Four nodule diameters were Fig. 1 Study design and workflow. MIP = maximum intensity projection, CAD1 = SyngoCT-CAD system (VA 20, Siemens Healthcare), CAD2 = Lung CAD system (IntelliSpace Portal DX Server, Philips Healthcare), CAD = computeraided detection. used: 5, 8, 10, and 12 mm. Lesion attenuation simulated solid and ground-glass nodules measuring 100 HU and 630 HU, respectively (Fig. 2). Solid and ground-glass nodules were randomly assigned to different bilateral lung segments and phantoms. In total, 133 ground-glass and 133 solid nodules were placed in 55 phantoms. The absolute number of nodules placed per phantom was 8 (range, 0 8). The allocation of nodules was recorded on a standardized spreadsheet by an independent investigator not involved in image interpretation. Each phantom was then scanned with both standard-dose and microdose CT on a 128-MDCT scanner (Somatom Definition Flash, Siemens Healthcare) featuring iterative reconstruction algorithms (IRIS, Siemens Healthcare) and an integrated detector system (Stellar, Siemens Healthcare). The following scan parameters were applied for the microdose CT scan: helical acquisition with 80 kvp and 6 mas. To achieve maximum dose reduction, automated dose reduction features were disabled on the scanner. A high pitch factor of 2.2 was chosen for further dose minimization. FOV was kept constant at 32 cm. Slice reconstruction thickness was 1 mm. Standard-dose chest CT scans were performed with 100 kvp and 100 Quality Reference mas. All other imaging parameters were the same used for the microdose scan. For raw dataset processing, level three iteration was applied; axial slices were reconstructed with three different kernels (i30, i50, and i70) at 1-mm slice thickness for interpretation; multiplanar reconstructions were not performed. Radiation dose was represented by the dose-length product (DLP) calculated by the scanner for a standardized body phantom with a diameter of 32 cm and a constant scan length (z-axis) of 35 cm to image the entire phantom. Effective dose for both the standard and microdose CTs was calculated by multiplying DLP by the organ-specific weighting factor published by the International Commission on Radiologic Protection [19]. Image Analysis After data acquisition and processing, images were transferred to the PACS server (PACS IDS7, Sectra). Simultaneously, the imaging data were sent to CAD1 and CAD2. Four independent readers with between 3 and 5 years experience in chest imaging read the MIP images on the PACS workstation. The readers were blinded to the location, distribution, composition, size, and number of embedded lung nodules. All scans were interpreted in a standard lung window setting (window center level, 600; window width, 1500) and with a MIP slab of 8 mm and a slice thickness of 2 mm. The image presentation sequence for each dose level (standard vs microdose) was randomly assigned. Furthermore, the radiologists were not aware of the CAD results at the time of MIP interpretation. They recorded the size and slice position at the maximum diameter for each suspected nodule. A fifth investigator, blinded to the distribution of the artificial lung nodules and the other readers MIP interpretation results, performed the automated CAD1 and CAD2 analysis. The true-positive, false-positive, and false-negative results for each reader and interpretation technique were calculated. Statistical Analysis Sensitivities for both CAD and MIP nodule detection on standard dose and microdose CT were calculated for each reconstruction kernel. The total amount of false-positive results as well as the false-positive rate in a phantom-based analysis were calculated and recorded. Dose, kernel, device, and nodule-based comparisons of sensitivities and false-positive rate were performed using the test of comparison of proportions (z-test). Because there were four readers for the MIP images, the mean sensitivities were pooled to 452 nodules. Artificial nodules placed and position recorded (truth standard) Standard-dose scan Reconstruction with i30, i50, and i70 kernels MIP (four readers) Readers mark nodule position on MIP Phantom Microdose Solid and ground-glass nodules (5-, 8-, 10-, 12-mm diameter) CAD1 and CAD2 (one reader) Reader records true-positive, false-positive, and false-negative CAD results AJR:207, August

3 Ebner et al. Dose levels were analyzed with the comparison of two means test. The level of significance was set to p < MedCalc (version , MedCalc Software) was used for statistical analysis. Results Nodule Detection With Computer-Aided Detection and Maximum Intensity Projection Nodule detection sensitivities of the two different CAD systems and MIP at varying dose levels are outlined in Table 1. At a standard dose, CAD1 showed similar overall detection sensitivities ranging from 96.0% to 97.3% (p > 0.6) for both solid and groundglass nodules at variable reconstruction kernels. Likewise, the best CAD performance on microdose CT was similar to that of standard-dose CT, with overall best sensitivities from each dose group of 96.0% and 97.3%, respectively (p = 0.61). In fact, no significant difference was seen in CAD1 performance between standard and microdose with one exception: The sensitivity for ground-glass nodules on microdose CT reconstructed with a hard lung window kernel (i70) was much lower when compared with standard-dose CT (78.8% vs 92.9%, respectively; p = ). CAD2 performance was more variable than that of CAD1. The sensitivities of CAD2 and CAD1 for solid lung nodules were overall comparable for standard and microdose scans (CAD1, %; CAD2, %). However, CAD2 sensitivities for ground-glass nodules were significantly lower than what was seen with CAD1. By way of comparison, the sensitivity for ground-glass nodules with microdose CT and an i50 kernel on CAD1 was 92%; CAD2 had a sensitivity of only 12.4% (p < , Table 1). Sensitivity range for all nodules, doses, and reconstruction kernels was % for CAD1 and % for CAD2. The corresponding false-positive rates are shown in Table 1. In the per-phantom analysis, falsepositive test results with both CAD systems were highest with microdose CT and softtissue kernels (CAD1, 8.3; CAD2, 64.7). In comparing both CAD systems, CAD2 delivered consistently higher false-positive rates relative to CAD1 (Table 1). Highest sensitivity for lung nodule detection was achieved with the combination of standard-dose CT, i50 kernel, and CAD1 (97.3%). Of note, a similar sensitivity reaching 96% was achieved with the combination of microdose CT using i30 or i50 kernels and CAD1. Using the MIP series as a stand-alone interpretive modality yielded sensitivity comparable with that of CAD1 for both solid and ground-glass nodules across variable dose and reconstruction kernels (Table 1). As was the case with CAD1, MIP series were superior to CAD2 in detecting ground-glass nodules. However, MIP interpretation resulted in significantly fewer false-positives relative to CAD1 in the phantom-based analysis (MIP range: per phantom; CAD1 range: per phantom; p < ). Comparing the best MIP result at standard dose level (99.1% sensitivity with i50 kernel) with the best CAD result (97.3% sensitivity on CAD1 with i50 kernel), there was no statistically significant difference in performance (p = 0.14). At microdose level, the best MIP sensitivity (97.6% with i30 kernel) was comparable with the best CAD sensitivity (96.0% with i30 kernel; p = 0.36). TABLE 1: Sensitivities and False-Positive Results for Computer-Aided Detection (CAD) Systems and Maximum-Intensity-Projection (MIP) Reconstructions System, Kernel, and Dose Solid Nodule Sensitivity (%) Ground-Glass Nodule Sensitivity (%) Cumulative Sensitivity (%) Total No. of False-Positive Detections No. of False-Positives per Phantom CAD1 i30 kernel Standard dose Microdose i50 kernel Standard dose Microdose i70 kernel Standard dose Microdose CAD2 i30 kernel Standard dose Microdose i50 kernel Standard dose Microdose i70 kernel Standard dose Microdose MIP i30 kernel Standard dose Microdose i50 kernel Standard dose Microdose i70 kernel Standard dose Microdose Note CAD1 = SyngoCT CAD system (VA 20, Siemens Healthcare), CAD2 = Lung CAD system (IntelliSpace Portal DX Server, Philips Healthcare). 284 AJR:207, August 2016

4 MIP and CAD Algorithms in CT for Lung Cancer Screening Performance of Microdose CT Mean phantom dose for microdose CT was significantly less than that for standard-dose CT ( msv vs 1.65 msv, respectively; p < ). Using the best MIP results for nodule detection, no statistically significant difference was seen in diagnostic accuracy and sensitivities between standard-dose and microdose CT (99.1% vs 97.6%, respectively; p = ). Likewise, the best CAD performance on microdose CT was similar to that of standard-dose CT, with overall best sensitivities from each dose group of 96.0% and 97.3%, respectively (p = 0.48). Different Combinations of Maximum Intensity Projection, Computer-Aided Detection, and Reconstruction Kernel MIP with standard-dose CT provided best performance for solid nodules with a soft-tissue reconstruction kernel (i30) compared with the medium (i50) and hard (i70) reconstruction algorithms (99.1% vs 97.4%, p < ). In the microdose examinations, i30 was also the most favorable reconstruction kernel for MIP interpretation when compared with i50 and i70 kernels (i30 vs i50, p < ; i30 vs i70, p < ). For the CAD analysis in microdose CT, i30 provided the best performance among the reconstruction algorithms (i30 vs i70, p < ). With respect to the CAD-kernel relation, no statistically significant differences were A Fig. 2 Anthropomorphic phantom with embedded nodules. A, Standard-dose CT (100 kvp, 100 mas, i70 kernel) with ground-glass (top image) and solid (bottom image) nodules (arrows). B, Microdose CT (80 kvp, 6 mas, i30 kernel) in same phantom with ground-glass (top image) and solid (bottom image) nodules (arrows). C, Maximum-intensity-projection reconstructions calculated from standard dose scan in A (top image, ground-glass nodule; bottom image, solid nodule). Arrows indicate nodules. B C AJR:207, August

5 Ebner et al. seen in sensitivity between the three kernel reconstructions (p > 0.8) within each CAD. There were also no significant differences in sensitivity between CAD systems and kernels for solid nodules; however, as previously noted, CAD2 performed significantly worse than CAD1 for ground-glass nodules at each kernel reconstruction level. Influence of Nodule Size and Composition The influence of nodule size on lung lesion sensitivity relative to CT dose and kernel reconstruction is shown in Table 2. In general, sensitivity increased with increasing nodule size for both CAD1 and MIP modalities with both standard and microdose CT and for all kernel reconstructions. The application of CAD1 performed as well for small nodules as it did for large nodules in standard and microdose technique (sensitivities between 5-mm and 12-mm nodules, p = ). Also, no statistically significant difference was seen in CAD1 performance for nodule detection between standard and microdose CT. In comparing CAD systems, CAD1 performed significantly better than CAD2 for all nodule sizes irrespective of dose or kernel. Of note, CAD2 sensitivities were inversely proportional to nodule size. The impact of kernel reconstruction on nodule detection is noted in Tables 1 and 2. For both CAD and MIP with standard-dose CT, differences in sensitivities between kernels were not statistically significant. However, when applied to microdose CT, kernel reconstructions had a more appreciable impact on nodule detection relative to size. For example, an i30 kernel was significantly better than an i70 kernel for detection of nodules that measured 5 mm when using MIP reconstructions for microdose CT (sensitivities of 94.8% vs 62.9%, respectively; p < ). No statistically significant difference was seen between kernels for larger nodules. Overall, MIP and CAD performed similarly (p = ) regardless of nodule size or CT dose protocol. MIP performance was independent of nodule composition (solid or ground-glass) for both standard and microdose CT (p = ). When operated with standard-dose CT, the CAD1 system showed unimpaired performance, regardless of nodule composition (p = 0.7). When tested with a hard lung kernel (i70) on microdose CT, CAD1 was significantly more sensitive for detecting solid nodules than for detecting ground-glass nodules (p < 0.042). CAD2 was less reliable for detecting ground-glass nodules regardless of dose and kernel employed (Table 1). Superior nodule perception was achieved when MIP and a soft-tissue kernel (i30) were used (Table 1). Satisfaction of Search in Maximum-Intensity- Projection Readings To address the relationship of the number of missed nodules and the number of nodules embedded in the phantom, we separately analyzed sensitivity on the basis of nodules per phantom (Table 3). A substantial satisfactionof-search effect was seen after readers identified the first nodule in phantoms equipped with two lesions. However, this effect did not apply in phantoms with more than two lesions (three to eight nodules, Table 3). Discussion Lung cancer is the leading cause of cancer-related deaths in the United States and worldwide [20]. Screening with low-dose CT has been shown to reduce lung cancer-specific mortality through early detection [2]. Low-dose CT screening for lung cancer has been endorsed by nearly every major international medical society and other agencies including the ACR, American Cancer Society, American College of Chest Physicians, Amer- TABLE 2: Size-Dependent Nodule Sensitivity for Computer-Aided Detection (CAD) Systems and Maximum-Intensity- Projection (MIP) Reconstructions Nodule Size i30 i50 i70 Standard Dose Microdose Standard Dose Microdose Standard Dose Microdose CAD1 5 mm a 87.5 ± ± ± ± ± ± mm a 98.2 ± ± ± ± ± ± mm b ± ± ± ± ± ± mm c 98.4 ± ± ± ± ± ± 2.3 CAD2 5 mm a 80.4 ± ± ± ± ± ± mm a 53.6 ± ± ± ± ± ± mm b 56.6 ± ± ± ± ± ± mm c 60.7 ± ± ± ± ± ± 6.4 MIP 5 mm a 94.8 ± ± ± ± ± ± mm a 96.5 ± ± ± ± ± ± mm b 98.1 ± ± ± ± ± ± mm c 98.4 ± ± ± ± ± ± 2.5 Note Values are mean ± SD (%). CAD1 = SyngoCT CAD system (VA 20, Siemens Healthcare), CAD2 = Lung CAD system (IntelliSpace Portal DX Server, Philips Healthcare). a n = 56. b n = 53. c n = AJR:207, August 2016

6 MIP and CAD Algorithms in CT for Lung Cancer Screening TABLE 3: Satisfaction of Search and Impact on Reader Performance No. of Nodules per Phantom Sensitivity (%) p vs Single-Nodule Phantom a NA b Note The number of images read of phantoms containing four nodules or fewer was 912 (pooled sensitivity, 92.2%). The number of images read of phantoms containing five or more nodules was 1812 (93.4%). The difference between the two groups was not statistically significant (p = 0.28). NA = not applicable. a Pooled from four readers and three kernels. b Statistically significant. ican Thoracic Society, National Comprehensive Cancer Network, U.S. Preventive Services Task Force, and Centers for Medicare & Medicaid Services. Growing acceptance of low-dose CT screening has increased interest among patients who are at high risk for lung cancer, their physicians, and insurers, many of whom offer coverage for annual low-dose CT screening examinations and anticipate growing use of this important service. Given the potential impact on a large patient population, low-dose CT screening must be properly performed and optimized. In this context, a discussion about the costs of CAD systems is warranted, because they may contribute to the overall cost-effectiveness of low-dose CT screening. The costs of CAD systems vary widely depending on the vendor and institutional contracts. Numerous large and small enterprises offer affordable solutions, with many scanner manufacturers offering CAD solutions embedded in their workspaces. With respect to effectiveness, CAD as a second reader increases sensitivity and reduces false-negatives at lower cost than double reading by a second radiologist [21, 22]. However, increased sensitivity with CAD comes at a cost of increased reading times. Beyer and colleagues [21] found a mean increase in reading time of 45 seconds when using CAD as a second reader compared with readings without CAD. Notwithstanding the increase in time for interpretation, current European lung cancer screening guidelines recommend the general use of CAD systems to improve quality, outcomes, and cost-effectiveness [22]. Many parameters influence the accuracy and efficiency of pulmonary nodule detection on chest CT. This study compared the detection rate of solid and ground-glass pulmonary nodules of variable size using different CT dose levels and reconstruction kernels using one of two different CAD systems or MIP processing for interpretation. The data show a significant advantage of CAD1 over CAD2, with the former exhibiting higher sensitivities in both the standard-dose and microdose spectra. With a significant difference in the detection of ground-glass lesions, CAD1 outperformed CAD2. Notably, readers using MIP alone showed better detection rates than those using CAD2. CAD2 performed best with standard-dose CT and a hard lung kernel (i70), which likely reflects the most common imaging parameters for routine clinical diagnostic chest CT, but the performance of CAD2 on lower dose examinations and datasets with softer reconstruction kernels was suboptimal. When we compared the performance of both CAD systems, an inverse relationship between nodule size and sensitivity for CAD2 was evident. Our findings indicate that CAD2 sensitivity decreases with increasing lesion diameter (Table 2). Since both CAD systems were tested with the same scans, this issue is most likely attributable to the software algorithm applied in CAD2. CAD detection sensitivity has been reported to typically decrease with decreasing nodule size [23, 24]. In a study by Yuan et al. [25], the CAD system tested missed seven large (> 10 mm) nodules out of 27 lesions. Yuan and colleagues were able to verify that these large nodules were missed because of immediate continuity with normal anatomic structures such as fissures, pleura, or lung vessels. We hypothesize that the CAD2 system used in this study suffered from similar segmentation errors. Radiation dose minimization places further demands on CAD systems. Systems cannot only perform well with standard-dose CT; current imaging guidelines and expectations necessitate increasing performance in a lowdose imaging range. In our study, CAD1 and the MIP reconstructions met these demands (Table 1). Comparing the sensitivity of the two CAD systems related to the image reconstruction kernel, microdose CT performed better in combination with the soft-tissue kernel (i30) on both systems. Inversely, the performance of both CADs applied on microdose CT decreased with i50 and i70 kernels. This trend is likely a result of variations in the signal-tonoise ratio. As a result of minimizing tube voltage and tube current in obtaining microdose CT, image noise increases and is directly related to the applied folding kernel with higher (harder) kernels producing inferior signalto-noise ratios compared with standard-dose CT. Using soft-tissue reconstruction kernels can improve the accuracy of nodule detection on microdose CT examinations when used in conjunction with CAD or MIP processing. In a study by Ebner et al. [6], readers favored softtissue kernels over standard lung kernels for microdose CT. In contrast, standard-dose chest CT was more accurate with a hard kernel (i70). The present data outline that soft-tissue kernels increase the sensitivity and consequently increase the false-positive rate in microdose examinations. Despite the influence of a satisfaction-of-search effect in phantoms containing two lesions, our findings suggest that MIP reconstructions might deliver similar sensitivities in a screening population and thus be able to serve as a substitute for CAD systems. Furthermore, MIP reconstructions are feasible without requiring more expensive CAD software solutions. The two CAD systems we studied exhibited distinct differences in their falsepositive rates in a phantom-based analysis. The false-positive results of CAD2 were more than 10 times higher than those of CAD1. Although CAD2 delivered markedly higher false-positive results, pooled sensitivity remained low. These discrepant results are not dose- or kernel-dependent and may be due to differences in software algorithms. Similar to the dose-kernel relationship, false-positive findings were minimized in standard-dose CT scans to which CAD1 was applied. The current study shares inherent limitations related to the nature of phantom experimentation, such as lack of anatomic detail, interindividual demographic variation, and imaging artifacts. These limitations were mini- AJR:207, August

7 mized as much as possible by using a lifelike replica of a human torso, but the results will require validation in prospective trials with human subjects. An additional consideration in using a phantom is that the artificial lung parenchyma and synthetic nodules result in images with sharp contrast compared with the airattenuating background, which may explain the excellent performance of all stand-alone reading devices, especially CAD1. Another point to consider is that the CAD software of both vendors is primarily designed to detect solid lung lesions. Although the vendors claim that their systems can also detect subsolid lesions, a lack of optimization for ground-glass nodules may explain the inferior performance of CAD2. Finally, a general limitation for CAD systems is the high false-positive rate (CAD1, maximum of eight false-positives per study; CAD2, maximum of 64.7 false-positives per study), which resulted in reduced specificity. Furthermore, we recognize that both dose and CT interpretation in a phantom study may not accurately reflect conditions encountered in clinical practice. The actual performance of CAD and MIP systems in a clinical setting may vary to a substantial degree not only because of the inherent differences of each system but also because of the nodule prevalence in a screening cohort. The present evaluation showed a high frequency of lung nodules (mean, 4.84 per patient); only five phantoms held no lesions, which may have introduced a selection bias. Conclusion The presented results suggest that lung cancer screening with microdose (submillisievert) CT is feasible for detection of solid and groundglass nodules of varying size and has potential for eventual implementation in clinical lung cancer screening programs. We also determined that microdose CT requires soft-tissue reconstruction kernels for both MIP and CAD systems to achieve acceptable sensitivities, whereas CAD applications in standarddose CT work most reliably with hard lung kernels. In our study, using a CAD system or MIP reconstructions best facilitated detection of solid and ground-glass nodules on CT. Because the CAD systems we evaluated did not perform similarly, we conclude that sensitivities of CAD algorithms can be improved with respect to nodule size and subsolid features. In settings where CAD is not available, MIP reconstructions remain a valuable adjunct to the Ebner et al. interpretation of chest CT scans for increasing sensitivity, with the advantage of significantly lower false-positive results. References 1. Boiselle PM. Computed tomography screening for lung cancer. JAMA 2013; 309: The National Lung Cancer Screening Trial Research Team. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011; 365: de Konig HJ, Meza R, Plevritis SK, et al. Benefits and harms of computed tomography lung cancer screening strategies: a comparative modeling study for the U.S. Preventive Services Task Force. Ann Intern Med 2014; 160: Roberts H, Walker-Dilks C, Sivjee K, et al. Screening high-risk populations for lung cancer: guideline recommendations. J Thorac Oncol 2013; 8: Frauenfelder T, Puhan MA, Lazor R, et al. Early detection of lung cancer: a statement from an expert panel of the Swiss university hospitals on lung cancer screening. Respiration 2014; 87: Ebner L, Bütikofer Y, Huber A. Lung nodule detection by microdose CT versus chest radiography (standard and dual-energy subtracted). AJR 2015; 204: Christe A, Szucs-Farkas Z, Huber A. Optimal dose levels in screening chest CT for unimpaired detection and volumetry of lung nodules, with and without computer assisted detection at minimal patient radiation. PLoS One 2013; 8:e Neroladaki A, Botsikas D, Boudabbous S, Becker CD, Montet X. Computed tomography of the chest with model-based iterative reconstruction using a radiation exposure similar to chest X-ray examination: preliminary observations. Eur Radiol 2013; 23: American College of Radiology website. ACR STR practice parameter for the performance and reporting of lung cancer screening thoracic computed tomography (CT). ACR/Documents/PGTS/guidelines/LungScreening. pdf. Accessed May 2, Mettler FA Jr, Huda W, Yoshizumi TT, Mahesh M. Effective doses in radiology and diagnostic nuclear medicine: a catalog. Radiology 2008; 248: Wall BF, Hart D. Revised radiation doses for typical x-ray examinations: report on a recent review of doses to patients from medical x-ray examinations in the UK by NRPB. Br J Radiol 1997; 70: Valencia R, Denecke T, Lehmkuhl L, Fischbach F, Felix R, Knollmann F. Value of axial and coronal maximum intensity projection (MIP) images in the detection of pulmonary nodules by multislice spiral CT: comparison with axial 1-mm and 5-mm slices. Eur Radiol 2006; 16: Godoy MC, Kim TJ, White CS, et al. Benefit of computer-aided detection analysis for the detection of subsolid and solid lung nodules on thinand thick-section CT. AJR 2013; 200: Jeon KN, Goo JM, Lee CH, et al. Computer-aided nodule detection and volumetry to reduce variability between radiologists in the interpretation of lung nodules at low-dose screening computed tomography. Invest Radiol 2012; 47: Roos JE, Paik D, Olsen D, et al. Computer-aided detection (CAD) of lung nodules in CT scans: radiologist performance and reading time with incremental CAD assistance. Eur Radiol 2010; 20: Choi WJ, Choi TS. Automated pulmonary nodule detection based on three-dimensional shapebased feature descriptor. Comput Methods Programs Biomed 2014; 113: Brown MS, Lo P, Goldin JG, et al. Toward clinically usable CAD for lung cancer screening with computed tomography. Eur Radiol 2014; 24: Christe A, Leidolt L, Huber A, et al. Lung cancer screening with CT: evaluation of radiologists and different computer assisted detection software (CAD) as first and second readers for lung nodule detection at different dose levels. Eur J Radiol 2013; 82:e873 e [No authors listed]. The 2007 recommendations of the International Commission on Radiological Protection: ICRP Publication 103. Ann ICRP 2007; 37: Siegel R, Ma J, Zou Z, et al. Cancer statistics, CA Cancer J Clin 2014; 64: Beyer F, Zierott L, Fallenberg EM, et al. Comparison of sensitivity and reading time for the use of computer-aided detection (CAD) of pulmonary nodules at MDCT as concurrent or second reader. Eur Radiol 2007; 17: Kauczor HU, Bonomo L, Gaga M, et al. ESR/ERS white paper on lung cancer screening. Eur Radiol 2015;25: Godoy MC, Cooperberg PL, Maizlin ZV, et al. Detection sensitivity of a commercial lung nodule CAD system in a series of pathologically proven lung cancers. J Thorac Imaging 2008; 23: Sahiner B, Chan HP, Hadjiiski LM, et al. Effect of CAD on radiologists detection of lung nodules on thoracic CT scans: analysis of an observer performance study by nodule size. Acad Radiol 2009; 16: Yuan R, Vos PM, Cooperberg PL. Computer-aided detection in screening CT for pulmonary nodules. AJR 2006; 186: FOR YOUR INFORMATION This article is available for CME and Self-Assessment (SA-CME) credit that satisfies Part II requirements for maintenance of certification (MOC). To access the examination for this article, follow the prompts associated with the online version of the article. 288 AJR:207, August 2016