Automatic calculation of patient size metrics in computed tomography: What level of computational accuracy do we need?

Abstract Objectives To compare the effectiveness of two different patient size metrics based on water equivalent diameter (D w), the mid‐scan water equivalent diameter D w_c, and the mean (average) water equivalent diameter in the imaged region, D w_ave, for automatic detection of accidental changes in computed tomography (CT) acquisition protocols. Methods Patient biometric data (height and weight) were available from a previous survey for 80 adult chest examinations, and 119 adult single‐acquisition chest–abdomen–pelvis (CAP) examinations for two 16 slice scanners (GE LightSpeed and Toshiba Aquilion RXL) equipped with automatic tube current modulation (ATCM). D w_c and D w_ave were calculated from the archived CT images. Size‐specific dose estimates (SSDE) were obtained from volume CT dose index (CTDI vol), using the conversion factors for a patient diameter of D w_c. Results CTDI vol and SSDE correlate better with D w_ave than with D w_c. R‐squared values of linear fits to CTDI vol of CAP examinations were 0.81–0.89 for D w_c and 0.93–0.94 for D w_ave (SSDE: 0.69–080 for D w_c, 0.87–0.92 for D w_ave). Percentage differences between D w_c and D w_ave were −4 ± 4% for chest and +5 ± 4% for CAP examinations (in % of D w_ave). However, small D w variations translated as larger variations in CTDI vol for these ATCM systems (e.g., a 24% increase in D w doubled CTDI vol). The dependence of CTDI vol on D w_ave was similar for chest and CAP examinations performed with similar ATCM parameters, while use of D w_c resulted in a clear separation of the same data according to examination type. Maximum D w variation in the imaged region was 5.6 ± 1.6 cm for chest and 6.5 ± 1.4 cm for CAP examinations. Conclusions D w_ave is a better metric than D w_c for binning similar‐sized patients in dose comparison studies, despite the additional computational effort required for its calculation Therefore, when implementing automatic determination of D w for SSDE calculations, automatic calculation of D w_ave should be considered.


| INTRODUCTION
Computed tomography (CT) is a powerful diagnostic tool, but CT imaging protocols should be optimized to minimize radiation exposure. Diagnostic reference levels (DRLs) have been a powerful tool in dose optimization, by establishing typical values of volume CT Dose Index (CTDI vol ) or dose-length product (DLP), for certain types of examinations performed on standard-sized (70 AE 3 kg) patients. [1][2][3][4] Modern CT scanners are equipped with automatic tube current modulation (ATCM), which adjusts tube current according to patient size and anatomical region, based on parameters set by the user. 5 The functioning of an ATCM system and the conditions to be set depend on the scanner manufacturer. For GE and Toshiba systems, three parameters must be specified: an "image quality index" related to image noise, and the minimum and maximum values of tube current, I min and I max . An adequate value of I max avoids unnecessary dose escalation in large or obese patients. I min is equally important to prevent excessive noise in smaller patients, particularly in low attenuation regions such as the lungs. 6,7 GE defines a parameter called noise index (NI) to specify the noise level, while Toshiba uses the standard deviation (SD). 8,9 ATCM systems have some limitations, and pediatric patients need separate imaging protocols, with different parameters (such as lower kV and lower I min ). The range of sizes in pediatric patients is immense, from babies to adolescents, and different protocols should be used according to child size/age. 8,10,11 Only adult patients will be considered in the present work.
Optimization of ATCM settings is a time-consuming process involving a radiologist who assesses image quality after each acquisition. This is usually done for only a few patients. If the results are considered acceptable, the protocol is implemented in a provisional fashion. Postacquisition assessment of examination doses and image quality is continued for some time, to confirm that settings are optimized for all patient sizes. 7 In this context, it is useful to have reference dose levels for different-sized adult patients. The American Association of Physicists in Medicine (AAPM) lists approximate reference values for different weight ranges in some typical examination protocols. 12,13 Reliance on ATCM systems increases the potential detriment of nonoptimized settings and accidental changes to previously optimized protocols, as follows. An accidental increase in target image noise will result in a degradation of image quality for all patients, which should be quickly detected by radiologists. However, a decrease in target image noise will increase examination doses and image quality for smaller adults, while scanner output for large patients is limited by I max . Likewise, an unnecessarily high value of I min increases examination doses for small adults, with no degradation of image quality. Both situations lead to saturation of the tube current (at I max or I min ) for an increased number of patients, 7 but this may go unnoticed under a heavy workload, or be dismissed through overconfidence in the automated system.
With the introduction of PACS (Picture Archiving and Communication Systems) in radiology, several vendors have developed radiation dose index monitoring (RDIM) software, which collects dosimetric information from imaging studies and stores it in a relational database. 14 RDIM systems are a powerful tool to identify accidental changes and outliers, and determine where optimization is needed. However, patient biometric data (height and weight) are not usually available in PACS. Therefore, an accidental change which affects only small adults is hard to recognize quickly, because individual examination doses are still in the expected range (e.g., a 50-kg adult imaged with a CTDI vol adequate for a 90-kg patient). Naturally, dosimetric data from thousands of examinations will include patients of all sizes and can be compared between different institutions and scanners. But this is population-dependent and also impractical for quick detection of changes and nonoptimized protocols.
The AAPM Task Group 204 proposed the use of size-specific dose estimates (SSDE) for patient dose comparisons. SSDE is an estimate of patient dose at the center of the imaged region, obtained from CTDI vol using conversion factors f(D eff ) related to the effective diameter of the patient, calculated from the measured anteroposterior (AP) and lateral (LAT) patient dimensions, D eff = √(AP•LAT). 15 To improve the calculation of SSDE by taking into account patient attenuation, the AAPM Task Group 220 proposed describing patient size in terms of water equivalent diameter (D w ). TG220 also suggested that D w could be calculated automatically by the CT scanner for all patients, with no user intervention, and the results stored in the DICOM header of CT images. 16 An automatically calculated D w would allow binning of similar-sized patients in RDIM databases for comparison of examination doses. SSDE values for adults may vary with patient size, depending on the ATCM system. 17 Leng et al. have shown that SSDE can be calculated with less than 10% error using the examination CTDI vol and the value of D w obtained from the mid-scan CT slice (D w_c ). 16 However, values of D w along the imaged region, D w (z), may be useful for estimating organ doses. 16,18 Obtaining D w (z) values requires longer computational times, but it also allows calculation of the mean value (average) of D w (z), D w_ave .
As the response of ATCM systems is based on patient attenuation, D w_ave is the quantity more closely related to the examination's mean CTDI vol. 19 Anam et al. recently reported on the implementation of automatic contouring for calculation of D w , and showed that D w_ave could be obtained with reasonable accuracy from only nine images, for head and thorax examinations. 20 This is still nine times the computational effort required for calculating D w_c . Differences between D w_c and D w_ave were found to be less than 10%, 20 which agrees well with data reported by other authors. 21 SARMENTO ET AL.

| 219
The aim of this study was to compare D w_c and D w_ave as patient size metrics, for the purpose of ATCM optimization and detection of accidental changes; and to determine whether the difference between the two is sufficient to justify the additional computational effort required to automatically determine D w_ave in addition to D w_c.
This study also assessed the interdependence of metrics and the feasibility of using D w metrics in nonautomated scenarios, for retrospective comparison with older data.

2.A | Data collection
This study took advantage of existing biometric data, which had been collected during a routine internal survey, after confirmation by the radiologists that image quality was satisfactory. Patient biometric data (height and weight) were available for 80 chest and 119 single- Both CT11 and CT14 are 16-slice scanners, equipped with ATCM in both the longitudinal direction (z-axis modulation) and the perpendicular plane (xy or angular modulation). The combination of these two is known as 3D modulation. Patients are randomly assigned to one scanner or the other, depending on equipment availability and internal logistics. Two orthogonal scout images (tube positions 0°and 90°) were acquired before each examination, in the order recommended by each manufacturer. The acquisition parameters are summarized in Table 3.
Proper functioning of the ATCM system and scanner indications of dosimetric parameters were checked at acceptance, and then annually, following the protocols and recommendations of the Spanish Medical Physics Society. 22 To reduce patient dose in CT examinations, it is important to limit anatomical coverage (scan range) to the area of clinical concern. 7 Appropriate restriction of anatomical coverage minimizes the scan length, whereas the optimization of ATCM parameters is reflected in the examination's mean CTDI vol . Both influence the dose-length product (DLP). In this work, the examination CTDI vol (mean CTDI vol for the 32 cm diameter CTDI phantom) was chosen as the dosimetric parameter of interest and obtained from the dose summary archived in PACS.

2.B | Retrospective data analysis using attenuation metrics
Reconstructed CT images can be used to calculate D w , provided the reconstruction kernel is linear and quantitative (not edge-enhancing or otherwise nonlinear). 16 The image series obtained with the SOFT (GE) and FC08 (Toshiba) reconstruction kernels were used for this study.
The field of view (FOV) used in chest and CAP examinations usually includes the outer contour of the patient. Visual observation of the CT images confirmed that, for the majority of the examinations, the whole contour of the skin was visible in the entire imaged region, except near the shoulders. Examinations where a large part of the patient's contour was outside the FOV were excluded from the dataset. These situations were too rare to justify a correction based on air border proportion, as suggested by Ikuta et al. 23 According to the AAPM Task Group 220, the water equivalent diameter (D w ) of an object is related to its water equivalent area 16 If <CT> ROI is the mean CT number in a ROI be determined from a CT image using eq. (1) 16 : The air surrounding the object should have negligible impact on the result. 16 To account for the attenuation of the CT to a spreadsheet, and D w (z) was calculated using eq. (2) 16 : The automated method to obtain D w was tested using two cylin-  Table 4 for linear fits to CTDI vol and SSDE data for CAP examinations.
Before the widespread use of ATCM systems, Menke tested different surrogates for mean patient attenuation and concluded that T A B L E 3 Acquisition parameters used in both scanners. the correlation between patient attenuation and body mass index (BMI) was no better than with patient weight. 19 A similar result was obtained in this study, as shown in Table 4.
The mean values and standard deviation (SD) of CTDI vol and SSDE are presented in Table 5, for the real examinations and for the simulated scenarios.

3.C | Interdependence of different metrics
The patient sample considered in this work is representative of a particular population of oncological patients (Tables 1 and 2).
Mean male and female heights agree well with known statistics for the Portuguese population in this age-group. 24 The data obtained in this work were compared with the earlier study by   For CAP examinations, the variation of D w_ave with patient weight is similar to that obtained by Menke for abdominal examinations (Fig. 5).
There is good correlation between D w_c and D w_ave, for both examination types studied, as shown in Fig. 6. The two metrics are very similar, with maximum percentage differences below 15% (in % of D w_ave ) as shown in Fig. 7 and summarized quantitatively in Table 6. There appears to be some separation of male and female patients, probably related to differences in body habitus. The mean D w_c À D w_ave difference for chest CT was À4 AE 4% (in % of D w_ave ), which is comparable with the À1 AE 4% reported by Anam et al. 20 The variation of D w found in each examination, D w _ max À D w _ min , is similar for male and female patients, as shown in Fig. 7 and  (Table 6).    More data are necessary, especially as automatic selection of tube voltage may soon be a widespread option as well. 28 The example presented here merely highlights the importance of choosing a patient size metric which reduces data dispersion to a minimum, to improve detection of normal trends and outliers.

4.C | Interdependence of different metrics
The comparisons shown in Fig. 5 are an encouraging result, suggesting that the study of a sufficiently large number of different populations and anatomical regions might provide a conversion between patient weight and D w_ave , for each examination type.
This would allow comparison of newer large-scale data based on D w metrics with the existing studies and standards based on patient weight.
As reported by other authors, the impact of D w_ave À D w_c differences on SSDE values is small, 21 because the two metrics have quite similar values. This is reflected in the small data dispersion seen in Fig. 3(b) and 3(c). However, the lower dispersion of dosimetric data when D w_ave is used as a metric for patient size suggests that the effect of small D w_ave À D w_c differences is probably ampli- T A B L E 6 Differences between D w_c À D w_ave and D w (max) À D w (min), represented as mean AE standard deviation, for different examinations and patient groups (see plots in Fig. 7). These results highlight the importance of automatic calculation of D w_ave , despite the additional computational effort involved.
Moreover, these data suggest that D w_ave needs to be determined with great accuracy and in a standardized manner, if it is to be used for comparisons between different CT scanners and different institutions. In this work, the mean difference between automated and manual D w was found to be 0.2 AE 1.2% (0.05 AE 0.32 cm), but the mean absolute difference was 1.0 AE 0.7% (0.26 AE 0.19 cm), with a maximum difference of 2.2% (0.6 cm).
Drawing a ROI including the entire FOV is computationally fast and simple. But the greater accuracy (maximum 0.5% difference) reported for automatic contouring 20 should prove useful for dose comparisons, particularly during initial studies and acquisition of baseline data.

| CONCLUSIONS
This study highlights the importance of automatic calculation of D w (z), not just for organ dose estimation as already recommended, 16,18 but also to make D w_ave available to end users as a patient size metric for binning similar-sized patients in RDIM systems.
Despite the small percentage difference between D w_c and D w_ave (À4 AE 4% for chest and +5 AE 4% for CAP examinations in this study), both CTDI vol and SSDE present a stronger correlation with D w_ave than they do with D w_c . Our data suggest D w_c values reflect localized anatomy characteristics. The lower dispersion of dosimetric data obtained with D w_ave makes it easier to identify trends and outliers. This is useful for ATCM optimization and detection of accidental changes. Use of D w_ave also reduces dependence on examination type, which may be difficult to identify accurately in large-scale databases. Therefore, when implementing automatic determination of D w_c for SSDE calculations, automatic calculation of D w_ave should definitely be considered as well, despite the additional computational effort involved.
Use of D w metrics is not yet widely implemented in CT scanners and RDIM systems, but it is important to acquire baseline data for D w_ave metrics and to establish comparisons with existing standards based on patient weight. Our experience shows that small-scale studies using D w_ave metrics are feasible in nonautomated scenarios and may be used initially to acquire baseline data from new and retrospective studies.