点击显示 收起
1 From the Institutes of Clinical Radiology (U.J.S., N.H., T.K.H., C.H., C.R.B., M.F.R.) and Medical Informatics, Biometry, and Epidemiology (A.C.), University of Munich, Germany. From the 2000 RSNA scientific assembly. Received November 14, 2000; revision requested December 26; final revision received July 25, 2001; accepted August 20.
| ABSTRACT |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|
MATERIALS AND METHODS: A multi-detector row spiral CT protocol for the diagnosis of pulmonary embolism was used that consisted of scanning the entire chest with 1-mm collimation within one breath hold. In 17 patients with central pulmonary embolism, the raw data were used to perform reconstructions with 1-mm, 2-mm, and 3-mm section thicknesses. For each set of images, each subsegmental artery was independently graded by three radiologists as open, containing emboli, or indeterminate.
RESULTS: For the rate of detection of emboli in subsegmental pulmonary arteries, use of the 1-mm section width yielded an average increase of 40% when compared with the use of 3-mm-thick sections (P < .001) and of 14% when compared with the use of 2-mm-thick sections (P = .001). With the use of 1-mm sections versus 3-mm sections, the number of indeterminate cases decreased by 70% (P = .001). Interrater agreement was substantially better with the use of 1-mm and 2-mm sections than with the use of 3-mm sections.
CONCLUSION: For the diagnosis of subsegmental pulmonary emboli at multi-detector row CT, the use of 1-mm section widths results in substantially higher detection rates and greater agreement between different readers than the use of thicker sections.
Index terms: Embolism, pulmonary, 564.813, 68.721 • Pulmonary arteries, CT, 564.12115
| INTRODUCTION |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|
Within the past several years, multi-detector row spiral CT has been introduced into clinical practice (9,10). The most prominent feature of multi-detector row spiral CT is its high acquisition speed. This increased speed can be used either to quickly cover large volumes, or, when narrow collimation is used, to increase spatial resolution and reduce partial volume averaging (11).
In a carefully designed study, Weg et al recently reconstructed identical thin-collimation multi-detector row spiral CT data sets of the liver with varying section thicknesses (12). These investigators were able to demonstrate that thin sections are superior to thicker sections in increasing the rate and the confidence of detection of small liver lesions (12). We adopted a similar strategy for our study, the purpose of which was to compare the efficacy of different reconstruction thicknesses of 1-mm collimation multi-detector row spiral CT data sets of the chest in the detection of subsegmental pulmonary emboli in patients with documented PE.
| MATERIALS AND METHODS |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|
Multi–Detector Row Spiral CT Scanning Protocol
Scanning was performed with a multi-detector row spiral CT scanner (VolumeZoom; Siemens Medical Solutions, Forchheim, Germany) with four detector arrays. Patients were scanned caudocranially within one breath hold. The entire thorax was included in the scan range. The scans were obtained with 4 x 1-mm collimation, with a table feed of 6 mm per 500-msec scanner rotation (12 mm/sec). This results in a pitch of 6, which is equivalent to a pitch of 1.5 in conventional CT systems. Scanning was performed at 120 kV and 120 mAs. Studies were enhanced with 120 mL of iopentol (Imagopaque 300; Nycomed-Amersham, Ismaning, Germany) injected at 4 mL/sec with an empiric scan delay of 16 seconds. With this protocol we consistently achieved high and uniform contrast enhancement throughout the thorax in all patients. For each patient, the raw data were retrospectively reconstructed with section thicknesses of 1, 2, and 3 mm, resulting in a total of 51 data sets. Two-millimeter and 3-mm sections were reconstructed by "fusing" the data from the original 1-mm spiral acquisition into thicker sections.
Image Evaluation
All images were evaluated on a monitor in "scroll-through" or cine mode at a PC-based workstation (Magic View 1000; Siemens Medical Solutions) with the standard window setting for viewing CT angiograms (width, 500 HU; center, 80 HU) at our institution. Images were independently evaluated by three experienced radiologists (U.J.S., N.H., T.K.H.) with 4, 9, and 13 years of experience, respectively, in evaluating thoracic CT scans. Readers were blinded to the patients’ names, to the section thickness used, and to each other’s ratings. Different reconstruction thicknesses of different scans were randomly mixed and were not presented in a preset order. To identify subsegmental arteries, the nomenclature outlined by Remy-Jardin et al (6) was used. A total of 42 subsegmental arteries are described in this nomenclature. Prior to the image analysis, a training session was held during which the readers agreed on the strategy for the analysis of the images. A subsegmental artery was considered to contain emboli when it manifested a definite filling defect on at least two consecutive sections. Each subsegmental artery in each patient was classified, on images created with each of the three section thicknesses, as open (ie, free of embolic material), containing emboli, or indeterminate. In the case of anatomic variants, the variant artery was designated according to the lung subsegment that it supplied. If a subsegmental artery could not be visualized, it was considered to be in the indeterminate category. The arteries of one lower lobe of the lung could not be analyzed in two patients due to the presence of a lung opacity in one patient and a pleural effusion in both patients. The arteries of these two lobes only were excluded from analysis.
Statistical Analysis
All analyses were performed with SAS 8.01 for Windows (SAS Institute, Cary, NC). Pair-wise agreement of the raters was assessed with simple kappa () statistics (13). was calculated separately for each section thickness. For the calculation of , the status of each vessel (ie, embolized, open, or indeterminate) was regarded as a nominal variable (ie, two ratings were regarded to be in agreement only if both readers chose the same category). For the purpose of comparing the number of emboli detected with each section thickness, an embolus was assumed to be present if all three raters had classified an artery as embolized. Statistical inferences were based on straightforward sign tests. Analogous analyses were performed to study the number of vessels that had been classified as indeterminate by at least one reader. We used the Bonferroni correction to account for multiple testing. We regarded P values of less than .012 as indicating a significant difference on a local level. Because the analyses included eight hypothesis tests, an overall level of .1 was preserved. In addition, the rate of detection of emboli in different areas of the lung with the three section thicknesses was compared, although the results of this comparison were not statistically tested.
| RESULTS |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|
|
|
|
|
Agreement among the three readers was highest for the 1-mm studies, with values of 0.73 (95% CI, 0.67–0.80), 0.78 (95% CI, 0.71–0.84), and 0.79 (95% CI, 0.73–0.85). values were similar for the 2-mm studies at 0.71 (95% CI, 0.65– 0.78), 0.74 (95% CI, 0.67–0.81), and 0.71 (95% CI, 0.65–0.78). Agreement was noticeably lower for the 3-mm studies, resulting in values of 0.38 (95% CI, 0.30–0.45), 0.42 (95% CI, 0.34–0.50), and 0.42 (95% CI, 0.34–0.50).
During the study, we observed that the higher rates of detection of emboli in subsegmental vessels with thinner sections were most substantial for the subsegmental vessels of the right middle lobe (RA4a, RA4b) and lingula (LA4a, LA4b, LA5a) and for some small vessels in the right lower lobe (RA7a, RA7b) (Figs 5, 6). These vessels, among others, tend to have a more oblique course relative to the transaxial scan plane (Figs 7–10). No noteworthy increase in the detection rate was observed for arteries with a more perpendicular course relative to the transaxial scan plane, such as RA1a, RA8a, RA8b, RA10b, LA1b, LA8a, and LA8b, although this was not statistically tested on an individual basis.
|
|
|
|
|
|
|
| DISCUSSION |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|
A 2-mm collimation single-detector CT scan with a pitch of 2 (ie, a 4-mm table feed per 0.75-second revolution), which has been proposed as the optimal acquisition protocol for single-section CT (6), allows coverage of a volume from the base of the heart to the bottom of the aortic arch (10–12 cm) within 19–23 seconds. A 1-mm collimation multi-detector row spiral CT acquisition with a pitch of 6 (ie, a 6-mm table feed per 0.5-second revolution), covers the same range in 8–10 seconds. Thus, despite the use of thinner, 1-mm collimation, the acquisition speed can be dramatically increased with multi-detector row spiral CT. Even if 1-mm collimation multi-detector row CT data are used to reconstruct thicker sections, as was performed in this study, image quality benefits because acquisition with narrow collimation of 4 x 1 mm effectively reduces partial volume effects and corresponding artifacts. The fusion of these data into thick sections restores low image noise. In addition, the slice-sensitivity profile of a thicker section reconstructed from 1-mm collimation multi-detector row CT data has sharper definition compared with a single-section spiral CT acquisition with comparable section thickness (22). The inclusion of the entire chest in a single breath hold with multi-detector row spiral CT, as in this study, enhances the diagnostic value of the examination because alternative or additional disease can also be detected in the apices and bases of the lung (23). Additionally, in cases without a known source of embolism, multi-detector row CT angiography of the pulmonary arteries can be performed in combination with multi-detector row CT venography of the lower extremities for a quick and comprehensive assessment of the pulmonary and venous embolus burden in patients suspected of having PE (24,25).
Although the advent of multi-detector row spiral CT has greatly expanded our diagnostic capabilities in patients suspected of having PE, these enhanced capabilities come in many cases at the price of additional radiation to the patient, especially when thin sections are used (11). With our scanning system, 1-mm collimation results in a 23% increase in effective radiation dose compared with 5-mm collimation when used in single-detector CT (22). Although we are convinced of the benefits of thin-section multi-detector row CT and advocate its use, especially in cases of suspected PE, these benefits must always be weighed against its increased radiation dose. Accordingly, heightened awareness of the additional radiation burden brought about by novel diagnostic techniques is warranted within the radiology community.
Our study design allowed us to evaluate multi-detector row spiral CT with different section widths in the detection of subsegmental emboli in patients with documented central embolic disease (Figs 1, 2). Because all data reconstructions were performed on the basis of the same 1-mm scan, unnecessary radiation to the patient was avoided and no bias due to patient motion or the use of different contrast materials was introduced. The results of our study suggest that the high spatial resolution conferred by 1-mm reconstructions of multi-detector row spiral CT data sets substantially improves the rate of detection of subsegmental pulmonary emboli compared with thicker section widths (Figs 3, 4). The increase in the detection rate observed in our study is most likely directly related to the accurate depiction, without volume averaging, of progressively smaller vessels with the use of thinner sections. For the same reason, the diagnostic confidence is improved, with substantially fewer subsegmental arteries deemed indeterminate on 1-mm and 2-mm sections compared with 3-mm sections (Fig 4).
This gain in diagnostic confidence with the use of thin sections also increased the reproducibility of findings between different observers, with notably better agreement between the three readers for 1-mm and 2-mm sections compared with 3-mm sections. Invasive pulmonary angiography for suspected acute PE has been discontinued in our institution and thus could not be used as a diagnostic standard. In recent reports, the interobserver agreement for diagnosing subsegmental pulmonary emboli on conventional pulmonary angiograms ranged from 45% (26) to 66% (27), which was considered unsatisfactory by the authors of both reports. Although these findings may not be directly comparable with our results, we believe that the high interobserver correlation for 1-mm and 2-mm sections in our study serves as a measure for the high degree of reproducibility that can be achieved with thin-section multi-detector row spiral CT in the diagnosis of small peripheral emboli.
The advantage of 1-mm sections over 2-mm and 3-mm sections was most noteworthy for subsegmental vessels with an anatomic course oblique to the scan plane (Figs 5, 6). Because of their small size and anatomic orientation, these vessels are often subject to volume averaging on scans that use thicker collimation and reconstruction widths. It is not always possible to differentiate hypoattenuation caused by volume averaging from that caused by embolic filling defects within an artery. Thus, confident exclusion of small emboli in such vessels is often not feasible. This may have contributed to the historically unsatisfactory performance of traditional CT protocols in the detection of emboli on the subsegmental arterial level (1,5). The improved visualization of oblique arteries that can be achieved with thin-collimation multi-detector row spiral CT (28) directly corresponds to the substantial increase in unanimously detected subsegmental emboli in our study (Figs 7–10).
We chose a patient population with central PE, because, even without an independent diagnostic standard, there was no doubt as to the absolute presence or absence of emboli in this group. At the same time, however, our focus on patients with known PE also limits our study, because in the presence of central PE a diagnosis of subsegmental emboli can be made with more confidence than in cases with a questionable peripheral filling defect in the absence of central emboli. Although our analysis is based on an extensive number of observations (ie, individual analysis of 660 subsegmental arteries), further limitations arise from the limited number of patients in our study group. Also, in the absence of an independent diagnostic standard, the accuracy of thin section multi-detector row spiral CT in the detection of subsegmental emboli ultimately remains uncertain. It is unclear whether invasive pulmonary angiography, given its known limitations, can provide suitable corroborative data. Better suited to this end may be studies that assess the respective performance of thin-section CT and invasive pulmonary angiography in correlation with a true independent diagnostic standard, like a recent study that used an animal model in which created emboli were subsequently identified and located after sacrifice of the animal (29). Such studies attest to the high accuracy of thin-section CT in the detection of small peripheral emboli (29).
On the basis of our experience, we believe that the high speed and spatial resolution that have become available with the advent of multi-detector row spiral CT improves our ability to noninvasively diagnose PE at all levels. For the diagnosis of subsegmental pulmonary emboli at thin-collimation multi-detector row spiral CT, the use of 1-mm section widths results in substantially higher detection rates than the use of thicker sections.
| STATISTICAL CONSULTANT COMMENTARY |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|
In the most typical situation, in which each of N subjects is classified into one of R categories by two classification methods, the observations may be summarized in an R x R contingency table, in which rows describe classification with one method and columns describe classification with the other method. If nij is the number of subjects classified into the row classification value i and the column classification value j, then one natural index of raw agreement is the proportion of subjects for which the two classification methods agree, according to the following equation:
The problem with po is that it reflects both chance agreement and agreement beyond chance. The fact that it reflects chance agreement can easily be seen in the following example: Assume the prevalence in a population of interest of characteristic A is 0.95. Furthermore, assume that one of two raters uses information to classify subjects as A or not-A. Note that if the other rater simply always diagnoses every patient as A, the two will agree with a po of 0.95. Therefore, a simple proportion-agreement score is insufficient to assess reliability.
The proportion agreement expected by chance, pe, is easily calculated from the marginal proportions of the two raters, exactly as in the 2 test of independence. So, to calculate a chance-corrected index of agreement, Cohen (Educ Psychol Meas 1960; 20:37–46) defined the index as follows:
The use of Cohen’s has been extended to differentially weight disagreements between ordered categories. For example, for a classification scheme with the categories none, mild, moderate, and severe, a weighted would score a disagreement between none and mild as less than a disagreement between none and severe. Use of the index has also been generalized to cases in which more than two classification schemes are used.
Note, however, that comparing across studies is problematic. The difficulty arises from the fact that depends on the marginal prevalence of the categories in the study. In general, agreement will be greater in studies with equal proportions of the categories. That is, for a fixed sensitivity and specificity, will be larger in a sample with 50% prevalence than in a sample with 10% prevalence.
| REFERENCES |
|---|
| Top ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION STATISTICAL CONSULTANT... REFERENCES |
|---|