Quantification of Distention in CT Colonography: Development and Validation of Three Computer Algorithms

¹ From the Department of Radiology, Stanford University Medical Center, MC 5105, 300 Pasteur Dr, Stanford, CA 94305 (P.W.H., D.S.P., S.N., R.B.J., J.M., A.J., C.F.B.); Mayo Medical School, Rochester, Minn (P.W.H.); Department of Radiology, University of California San Francisco and San Francisco Veterans Administration Medical Center (J.Y., A.S.G.); and Department of Radiology, New York University Medical School, NY (A.J.). From the 2000 RSNA scientific assembly. Received March 14, 2001; revision requested April 18; revision received July 9; accepted August 2. Supported in part by National Institutes of Health grant RO1 CA72023 and an award from the Society of Computed Body Tomography and Magnetic Resonance. 

	ABSTRACT

Top ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Three bowel distention–measuring algorithms for use at computed tomographic (CT) colonography were developed, validated in phantoms, and applied to a human CT colonographic data set. The three algorithms are the cross-sectional area method, the moving spheres method, and the segmental volume method. Each algorithm effectively quantified distention, but accuracy varied between methods. Clinical feasibility was demonstrated. Depending on the desired spatial resolution and accuracy, each algorithm can quantitatively depict colonic diameter in CT colonography.

　

Index terms: Colon, CT, 75.12115 • Computers, diagnostic aid, 75.12115 • Computed tomography (CT), computer programs, 75.12115 • Computed tomography (CT), image processing, 75.12115 • Phantoms

	INTRODUCTION

Top ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Colon cancer is the second leading cause of cancer death in the United States, with an estimated 56,300 deaths and 130,200 newly detected cases in the year 2000 (1). Computed tomographic (CT) colonography is a technique that shows promise in detecting precancerous adenomatous polyps (2–4). Previous studies have focused primarily on polyp detection sensitivity and computer methods for viewing the large data sets (2,3,5–7). An optimal CT colonographic study requires a clean well–air-distended colon (8–12) to prevent mucosal surfaces from collapsing together, thereby obscuring true lesions or creating pseudolesions (8–10). To allow retained fluid to move to dependent portions of the colon and optimize distention of various segments, most investigators now image the entire colon with the patient in supine and prone positions (9,13). On the basis of experience with barium enema examinations (14), it was also inferred that colonic distention in CT colonography would be improved by using bowel relaxants such as glucagon or butylscopolamine (Buscopan; Boehringer, Ingleheim, Germany). However, an initial subjective study comparing distention with and without glucagon showed no significant increase in distention or polyp detection sensitivity in glucagon-treated subjects (15,16).

Given the key influence of distention on CT colonography, we believed it would be beneficial to develop objective methods to quantify distention in the form of semiautomated computer software algorithms. Such algorithms have the potential to be used to catalog normal colon diameters, to study different bowel preparation and insufflation regimens, and to estimate how much of the colon surface is visible for polyp detection. Before systematic application to patient CT colonographic data, characterization of the performance of these algorithms is necessary. Thus, the purpose of our study was to develop algorithms that semiautomatically measure bowel distention in CT colonography, provide analytic validation in phantoms, and perform a preliminary application to a human CT colonographic data set to demonstrate clinical feasibility.

	Materials and Methods

Top ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Three algorithms were developed and applied to CT scans of physical colon phantoms and to computer-generated CT phantoms that mimicked curvatures, strictures, and haustral patterns found in the human colon. The algorithms were also applied to human CT colonographic data.

Imaging Data
Physical phantom.—To represent the colon with various cross-sectional diameters to simulate different degrees of distention, three physical phantoms were built, which each contained three 305-mm-long straight smooth acrylic tubes. Tubes had inner and outer diameters, respectively, of 70 and 78 mm, 44.5 and 51 mm, and 22 and 26 mm and were made watertight with plastic caps. Acrylic panels secured the tubes into a rectangular plastic container that served as a water bath during CT.

Each of the three physical phantoms was scanned at 0°, 45°, and 90° relative to the gantry with a single–detector-row helical CT scanner (HiSpeed Advantage; GE Medical Systems, Milwaukee, Wis) at 120 kVp, 150 mAs, with 3-mm collimation, pitch of 2, 1.5-mm image reconstruction interval, standard kernel, and a 25.6-cm field of view. These acquisition parameters are similar to those used in clinical CT colonography with the exception that a reconstruction field of view of 32–40 cm is used clinically.

Computer-generated phantoms.—To survey a wide range of geometric configurations, simulated colon phantoms were created by using CT simulator software (17). This program simulates the entire CT acquisition and reconstruction process, including partial volume effects, structured CT noise, and reconstruction artifacts by simulating forward projection and attenuation of x rays, noise, and filtered backprojection with helical interpolation. The simulator allows the user to create objects, define their shapes from geometric primitives such as ellipsoids, spheres, and cylinders, and to specify the attenuation values of the objects. The simulations were performed with a workstation (O2 [with an R12000 central processing unit and 512-MB main memory]; SGI, Mountain View, Calif).

Phantoms that represented a number of geometric configurations were simulated, including (a) three straight smooth cylinders with lengths and diameters identical to those of the physical phantom; (b) four straight cylinders (length, 300 mm; inner diameter, 60 mm; outer diameter, 80 mm), each with a single stricture of varying dimensions: 6 x 10 mm (diameter x length), 6 x 50 mm, 30 x 10 mm, or 30 x 50 mm; (c) one straight cylinder (length, 450 mm; inner diameter, 70 mm; outer diameter, 80 mm) with three 50-mm-long strictures and 10, 30, or 50 mm in diameter ; (d) two curved cylinders (inner diameter, 40 mm; outer diameter, 60 mm), both consisting of single semicircular curves: one wide curve approximately 770 mm long with radius of curvature of 150 mm and one hairpin curve approximately 630 mm long with radius of curvature of 37 mm; and (e) one straight cylinder (length, 460 mm; inner diameter, 50 mm; outer diameter, 60 mm) with simulated haustral folds spaced 20 mm apart .

fig.ommitted  Figure 1. Three-dimensional surface display rendering of a phantom with multiple strictures. Cut-away model shows inner aspect of phantom. Maximum internal diameter is 70 mm, and overall length is 450 mm. Strictures (arrows) reduce the diameter to 50 (A), 30 (B), and 10 (C) mm in the 50-mm-long segments.

 

fig.ommitted    Figure 2. Three-dimensional surface display rendering of a phantom with simulated haustral folds. Cut-away rendering shows full simulated haustra (black arrow) and approximately half of the opposite haustra (white arrow) at each level.

 
Each simulated phantom was modeled to contain an air-filled lumen (-1,000 HU) with a fat-density wall ( -100 HU) and a surrounding water bath (0 HU). Fat attenuation was chosen as the primary interface for the luminal air, as in our experience the well-distended colon wall is quite thin; therefore, surrounding abdominal fat is often the primary boundary. In any case, whether fat or soft-tissue attenuation was chosen as the air interface, no change in behavior of the computer algorithms was found during preliminary testing; the important factor is the large gradient in attenuation (1,000 HU) between the air and the boundary structure.

Each simulated phantom was scanned at 0°, 45°, and 90° with respect to the gantry. In addition, the haustral phantom was scanned three times in each orientation (nine total scans) to evaluate for reproducibility. CT acquisition and reconstruction parameters were set to mimic those used in clinical CT colonography, including 3-mm section thickness, pitch of 2, reconstruction interval of 1 mm, and tube current settings appropriate to generate noise values found in clinical CT colonography.

In summary, a total of 11 phantoms were simulated, which yielded 39 CT data sets for analysis.

Human CT colonographic data.—Each algorithm was applied to the supine-position CT colonographic data in a 48-year-old man who underwent imaging as part of an ongoing clinical trial. The subject’s colon was well prepared in terms of minimal retained fluid and only two short segments of poor distention. Informed consent was obtained, and the study was approved by our institutional panel for research in human subjects. After insufflation of the colon with room air, imaging was performed with a four–detector-row CT scanner (LightSpeed QXi; GE Medical Systems) with the following parameters: 2.5-mm section thickness, table speed of 7.5 mm per rotation (pitch of 3), 120 kVp, 56 mAs, and reconstruction interval of 1.25 mm.

Initial Image Processing
To apply the distention algorithms, preprocessing of the CT data was required in the form of segmentation and determination of a medial or centerline path through the lumen. The software used to accomplish these tasks was developed to navigate through the colon with a virtual camera (virtual endoscopy) and to create reformatted tomograms perpendicular or parallel to a centerline (18). For each CT data set (nine physical phantom scans, 39 simulated phantom scans), segmentation of the air-containing lumen was performed with a threshold of -700 HU. Next, a centerline path was generated through each of the segmented objects by specifying start and end points for each object, followed by iterative thinning. To minimize local fluctuations (zigzag) in the path that arose from the algorithm itself and from image noise, the centerline path was automatically smoothed over 25-mm segments by using a Hanning window (18). The path generation process is highly reproducible and has been reliably used to find the medial axis of other phantoms and structures in clinical CT data (19). The final centerline paths contain one point that corresponds to each image voxel traversed. Given the voxel dimensions of our data sets, path points were spaced approximately 1 mm apart for each of the test phantoms and for the clinical CT colonographic data. Taking all the paths generated on the phantoms together, there were approximately 12,400 path locations with which to test the distention algorithms. The human colon path was 1,520 mm long.

Computer Algorithms
Three algorithms were developed. On the basis of their modes of action, they were named as follows: (a) the cross-sectional area (CSA) method, (b) the moving spheres (MS) method, and (c) the segmental volume (SV) method. Each of these methods requires segmentation of the colonic lumen and a centerline path.  illustrates the concepts behind each algorithm. Detailed descriptions are in the Appendix.

fig.ommitted    Figure 3. Schematic of colon depicts the distention quantification algorithms with central path (dashed line). CSA method is shown by lines 1, 2, and 3. Note that plane 3 would show spurious CSA owing to colonic curvature, which causes the plane to intersect along the transverse colon (*) rather than at the opposite side of the flexure. MS method is shown by spheres 4 and 5. The larger sphere 4 depicts De averaged over a longer colon segment than does the smaller sphere 5. Spheres of a selected size are sequentially moved along the path to calculate De at points 6, 7, and 8. SV method is shown by regions 9, 10, and 11. In this method, the end planes are wobbled about the path point to obtain the minimum local CSA. The volume of the colon between the end planes is determined on the basis of colon segmentation.

 
CSA method.—An effective cross-sectional diameter of the approximately circular colonic lumen was derived on the basis of tomographic planes oriented perpendicular to the colonic axis. A plane perpendicular to the centerline was computed at each path point, and the intersection of this plane with the segmented colonic volume defined the local CSA of the colon. For the most accurate measurement of CSA, the method took into account partial volume averaging of voxels at the boundary zone between air and the bowel wall (Appendix). The effective colonic diameter (D_e), was calculated from the CSA. The CSA algorithm was applied to each set of CT data. Preliminary analyses with the CSA method revealed relatively noisy graphic depictions of colonic diameter (shown in Results) and that its accuracy was affected by imperfections in the centerline, particularly at areas of high luminal curvature ( line 3). For these reasons, we developed two further algorithms that determine an effective diameter by averaging over multiple path points.

MS method.—To create a moving average of D_e, the MS method generates a sphere centered at each path point, then computes the volume of intersection between the sphere and the local colonic segmentation (, details in Appendix). Sphere size is an adjustable parameter set by the user before the MS method is performed. To provide an accurate estimate of D_e, the sphere diameter must be larger than that of the colonic lumen. If the diameter of the sphere is smaller than that of the colon, the MS method may underestimate D_e (, sphere 5). To evaluate the effect of sphere diameters on distention measurements, the MS method was performed four times separately with each of the phantoms, with sphere sizes of 1.25, 1.5, 1.75, and 2.0 R, where R is the known maximum cross-sectional radius of the phantom.

SV method.—The SV method was developed to allow the flexible measurement of colonic distention at local, regional, and global levels. The algorithm divides the colon into subsegments by splitting the segmentation along cut planes located at user-defined path points , details in Appendix). To reduce errors that occur owing to oblique cut planes at the ends of the colonic segments, a subroutine was implemented to minimize the local colonic CSA for that plane. Colonic volume between each pair of cut planes was determined by counting voxels. This SV was divided by the segmental path distance, which resulted in an average effective CSA. Area was converted to an effective colonic diameter by using the formula for area of a circle. The main adjustable parameter in the SV method is segment length. To demonstrate the averaging effect of values for this parameter, the SV method was applied to each phantom with use of segment lengths of 25, 60, and 100 mm, and with the length set at one-half the total path length.

Preprocessing Computation Times
Preprocessing time with the human colon data set, including segmentation and centerline creation, was approximately 20 minutes. The time required to perform the CSA and SV methods was approximately 15 minutes, while the more computationally intensive MS method required approximately 6 hours to complete. Computation times were shorter for the phantoms owing to their shorter overall lengths compared with the human colon.

Data Analysis
Experimental results with each algorithm were compared with the corresponding true CSAs (for the CSA and MS methods) or volumes (for the SV method), which were calculated on the basis of the known geometry of physical or simulated phantoms. For the more complex geometry of the haustral phantoms, measurement tools in a professional three-dimensional modeling software package (RHINO, version 1.1; McNeel and Associates, Seattle, Wash) were used to determine truth values. Differences were measured between experimental points and truth values at each path point with the CSA and MS methods and for each segment with the SV method.

Descriptive statistics were used in data analysis. Metrics used to compare experimental results with truth and with each other included maximum, minimum, and root-mean-square (RMS) errors (in millimeters). RMS error was chosen because it is a well-accepted metric of typical measurement error, with increased sensitivity to large errors compared with mean absolute error. In addition, the percentage error was calculated by dividing the mean error generated by each algorithm by the mean diameter of each phantom.

	Results

Top ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
CSA Method
 shows the results of the CSA method applied to a scan of the multiple stricture phantom, while  shows the results of the CSA method applied to a scan of the haustral phantom. Both graphs show experimental effective diameter (D_e) and true effective diameter plotted along the length of the centerline path. Measurements at each point along the path allowed a high-spatial-resolution depiction of diameter. The experimental measurements were generally accurate and reflected the topography of all the phantoms studied. However, because the CSA method derives diameter measurements from the area of tomograms generated perpendicular to the path, the method generates spuriously large local diameters if there are fluctuations in the path so the perpendicular tomograms are perpendicular to the path but not to the object being quantified. This may occur if the object is curving, as can occur in the human colon , line 3). In addition, irregularities in the centerline path that occurred because of imperfect smoothing or at the ends of objects may generate cross-sectional planes that are not perpendicular to the object. In , abnormally large effective diameter measurements occurred in several areas. Other spurious diameters were measured in the hairpin phantom at 0° rotation owing to a local zigzag in the path at the apex of the curve.

fig.ommitted  Figure 4a. (a) Graph depicts data for the CSA method in a multiple-stricture phantom and experimentally determined De  as a function of path distance. Results closely parallel mathematic truth (solid line). Minor errors in De occurred between 120 and 180 mm and at path points above 400 mm  In these areas, local zigzag of the path caused the CSA planes to be slightly oblique to the phantom, which increased De. Data for the phantom were scanned at 45° to the gantry and were representative of results at 0° and 90° to the gantry. (b) Graph depicts data for the CSA method in a haustral phantom and experimentally determined diameter (De, thick line) from CSA method compared with mathematic truth (thin line). General shape of curves match over majority of haustral folds. Fine sampling of the phantom with this algorithm reproduces structural details of the phantom, though rapid fluctuations in diameter at haustral folds were not represented as accurately in their magnitude as their periodicity. Spuriously large De occurred in the vicinity of the first simulated fold (90-mm path distance) and at the end of the phantom (*) owing to imperfections in the central path. Data for the phantom were scanned at 45° to gantry and were representative of results at 0° and 90° to the gantry.

 

fig.ommitted  Figure 4b. (a) Graph depicts data for the CSA method in a multiple-stricture phantom and experimentally determined De  as a function of path distance. Results closely parallel mathematic truth (solid line). Minor errors in De occurred between 120 and 180 mm and at path points above 400 mm (*). In these areas, local zigzag of the path caused the CSA planes to be slightly oblique to the phantom, which increased De. Data for the phantom were scanned at 45° to the gantry and were representative of results at 0° and 90° to the gantry. (b) Graph depicts data for the CSA method in a haustral phantom and experimentally determined diameter (De, thick line) from CSA method compared with mathematic truth (thin line). General shape of curves match over majority of haustral folds. Fine sampling of the phantom with this algorithm reproduces structural details of the phantom, though rapid fluctuations in diameter at haustral folds were not represented as accurately in their magnitude as their periodicity. Spuriously large De occurred in the vicinity of the first simulated fold (90-mm path distance) and at the end of the phantom (*) owing to imperfections in the central path. Data for the phantom were scanned at 45° to gantry and were representative of results at 0° and 90° to the gantry.

 
Quantitation of diameter errors for the CSA method is shown in . The maximum and minimum errors represent the largest or smallest errors along the entire path for the phantom. The RMS error represents the average error along the path and shows the overall accuracy of the method. The average error is also represented in relative terms by the mean percentage error. Overall, the CSA method was accurate, with a mean RMS error of 2.01 mm, mean percentage error of 2.50%, and mean minimum error of 0.00590 mm. As shown in the maximum error column, however, the sporadic errors due to path imperfections or object curvature resulted in substantial local errors at some position in almost all of the phantoms, as shown by the maximum error values (mean, 20.1 mm).

fig.ommitted  TABLE 1. CSA Method Measurement Error in Various Colon Phantoms

 
MS Method
 shows representative data for the MS method for the multiple stricture phantom  and the haustral phantom . This algorithm produced smooth curves that tracked the major trends in distention (eg, strictures) but that tended to smooth over fine structural details (eg, simulated haustra). Smoothing is an inherent characteristic of the MS method because the algorithm calculates an average diameter over a segment of colon contained within a local sphere centered at each path point.

fig.ommitted  Figure 5a. (a) Graph depicts data for the MS method in a multiple-stricture phantom. The solid line shows the true diameter of the phantom. Experimental results for De are for spheres of different sizes, where R is the largest radius of the phantom colon: 1.25 , 1.5 , 1.75 , and 2.0  R. While results for each sphere size show variations in De owing to strictures, accuracy in depicting true diameter was lower for larger spheres (eg, 2.0 R) than for smaller spheres. Data for phantom were scanned at 45° to the gantry and were representative of results at 0° and 90° to the gantry. (b) Graph depicts data for the MS method in a haustral phantom. Stacked plots show De compared with true diameter (solid lines) for spheres of various diameters: 1.25 , 1.5 , 1.75 , and 2.0  R. Since the method averages over a segment of colon proportional to the sphere size, detail of individual haustral folds is not resolved. At the ends of the phantom, De begins below the true diameter and gradually approaches truth (see explanation in text).

 

fig.ommitted  Figure 5b. (a) Graph depicts data for the MS method in a multiple-stricture phantom. The solid line shows the true diameter of the phantom. Experimental results for De are for spheres of different sizes, where R is the largest radius of the phantom colon: 1.25 , 1.5 , 1.75 , and 2.0  R. While results for each sphere size show variations in De owing to strictures, accuracy in depicting true diameter was lower for larger spheres (eg, 2.0 R) than for smaller spheres. Data for phantom were scanned at 45° to the gantry and were representative of results at 0° and 90° to the gantry. (b) Graph depicts data for the MS method in a haustral phantom. Stacked plots show De compared with true diameter (solid lines) for spheres of various diameters: 1.25 , 1.5 , 1.75 , and 2.0  R. Since the method averages over a segment of colon proportional to the sphere size, detail of individual haustral folds is not resolved. At the ends of the phantom, De begins below the true diameter and gradually approaches truth (see explanation in text).

 
Error values for the MS method with sphere size of 1.5 R are shown in  (other sphere sizes yielded similar results).  summarizes and compares error with various sphere sizes. Overall, D_e measurements with the MS method were accurate to within 6.1–7.6 mm, on the basis of RMS error (16.5%–22.7%, mean percentage error), depending on the sphere size. The averaging effect with the MS method resulted in substantial local measurement errors in areas where there was a sudden change in distention. RMS error and mean absolute percentage error were largest in those phantoms with the most variation in distention (phantoms that contained multiple or severe strictures, such as multiple strictures and stricture 2 in ). In addition, as shown in , overall RMS error increased as sphere size increased, which illustrates that the degree of smoothing and insensitivity to changes in distention were proportional to the size of the sphere. Minimum and maximum errors, however, were largely unaffected by sphere size.

fig.ommitted  TABLE 2. MS Method Measurement Error

 
Another feature of the MS method is that distention was systematically underestimated at both ends of the phantoms . This can be understood by considering two cases, the first with a sphere centered at the first path point and the second with a sphere centered at a midpath point. Because no air-containing voxels exist before the beginning of the path, the intersection of the first sphere with the segmented voxels yields a smaller colonic volume, and thus D_e, than the second sphere. The same holds true at the termination of the path. This behavior generated essentially all of the MS method measurement error observed in the phantoms with constant distention throughout their length (ie, the smooth straight and curved cylinders). While the errors generated at the ends of the paths could potentially be corrected, thereby improving the results, we have not yet modified the algorithm to perform such correction.

SV Method
 shows results of the SV method applied to a scan of the multiple stricture phantom, and  shows the SV method measurements from a scan of the haustral phantom. Because the SV method makes one volume measurement for each user-defined segment and calculates D_e on the basis of this measurement, this method results in a global average of distention over the entire selected segment. Thus, sensitivity to structural detail and changes in local distention decreases as segment length increases. As shown in , measurements made with the SV method with 25-mm-long segments began to resolve variations in distention caused by simulated haustra, whereas the SV method produced a much more global view of distention with segments that were one-half the length of the path.

fig.ommitted  Figure 6a. (a) Graph depicts data for the SV method in a multiple-stricture phantom. Stacked plots show De compared with true diameter for volume segments of various lengths: one-half total phantom length  and total phantom length of 100 , 60 , and 25  mm. In the top plot, the phantom was divided into only two segments, so De was averaged over several strictures. Dividing the phantom into smaller volume segments resulted in higher fidelity sampling of true morphology, which is best seen in the bottom plot. Since partitions between the volume segments were not necessarily aligned with the strictured zones of the phantom, De represents a weighted average over portions of the phantom with both full diameter and stricture depending on coincidence of the strictures and volume segment partitions. Data for phantom were scanned at 0° to the gantry and are representative of results at 45° and 90° to gantry. (b) SV method in haustral phantom. Stacked plots show De compared with true diameter (solid lines) for volume segments of various lengths: one-half total phantom length  and total phantom length of 100 , 60 , and 25  mm. Since the algorithm computes mean De over the volume segment, structural detail of the simulated haustra is not depicted. Smaller volume segments showed more variation in De along the path.

 

fig.ommitted  Figure 6b. (a) Graph depicts data for the SV method in a multiple-stricture phantom. Stacked plots show De compared with true diameter for volume segments of various lengths: one-half total phantom length  and total phantom length of 100 , 60 , and 25  mm. In the top plot, the phantom was divided into only two segments, so De was averaged over several strictures. Dividing the phantom into smaller volume segments resulted in higher fidelity sampling of true morphology, which is best seen in the bottom plot. Since partitions between the volume segments were not necessarily aligned with the strictured zones of the phantom, De represents a weighted average over portions of the phantom with both full diameter and stricture depending on coincidence of the strictures and volume segment partitions. Data for phantom were scanned at 0° to the gantry and are representative of results at 45° and 90° to gantry. (b) SV method in haustral phantom. Stacked plots show De compared with true diameter (solid lines) for volume segments of various lengths: one-half total phantom length  and total phantom length of 100 , 60 , and 25  mm. Since the algorithm computes mean De over the volume segment, structural detail of the simulated haustra is not depicted. Smaller volume segments showed more variation in De along the path.

 
Representative error measurements for the SV method are listed in . For all physical and simulated phantoms studied, the SV method measurements were on average accurate to within 0.808–1.97 mm on the basis of RMS error (1.67%–4.66%, mean percentage error). Also, on average, maximum error generally decreased as segment length increased .

fig.ommitted  TABLE 3. SV Method Measurement Error

 
Reproducibility of Results
 shows reproducibility data for all three algorithms applied to nine simulations of the haustral phantom, three in each orientation (0°, 45°, and 90°). Interscan variability may arise owing to scan acquisition or image processing factors. Reproducibility was excellent in general, however, as evidenced by small SDs for maximum and minimum errors and RMS error.

fig.ommitted  TABLE 4. Reproducibility of Algorithms in Haustral Phantom

 
Comparison of Distention Measurement Methods
Comparison of the error values for all three methods from  yields several observations. The mean RMS and mean errors were largest for the MS method, with increasing sphere size generating larger errors. This reflects the fact that the MS method calculates a mean value for local distention. RMS error for the MS method was 3.0–3.8 times greater than that for the CSA method. In turn, the RMS error with the CSA method was 2.5 times that of the SV method for the longest segments. Decreasing the segment length in the SV method generated larger errors; with 25-mm-long segments, RMS error approached that of the CSA method. Comparison of the range and distribution of minimum and maximum errors showed the most variability in measurement error with the CSA method and the least with the SV method. Mean maximum error was similarly high for the CSA and MS methods while substantially lower for the SV method.

Human Colon CT Colonographic Data
Initial feasibility of applying the distention algorithms to CT data in a human colon is shown in  and .  shows a three-dimensional rendering of the colonic air column, and shows results with each of the distention methods. The CSA method measurementswere characterized by a high spatial resolution but noisy representation of distention and by sharp upward spikes in D_e at points where high colonic curvature resulted in CSA planes instantaneously perpendicular to the centerline path but oblique to the colonic axis. Areas of nondistention were accurately represented by D_e values of zero. The MS method  produced a noticeably smoother curve than did the CSA method owing to local averaging. Finally, the SV method algorithm  yielded less detail but depicted segmental distention adequately. Owing to averaging, however, the two areas of local collapse were not evident.

fig.ommitted  Figure 7. Initial clinical CT colonographic application of distention algorithms. Three-dimensional point cloud representation was obtained with segmentation of colonic air at a threshold of -850 HU. Centerline path is shown as a white line through the center of the point cloud. The largest diameters occur in the rectum (R) and cecum (C). Two collapsed portions (*) of descending colon are shown.

 

fig.ommitted  Figure 8a. Distention analysis in human colon. (a) Graph depicts data with the CSA method. Distention profile shows undulating baseline, which represents De from the rectum (point 0) to cecum (point 1520). High-frequency spikes (*) in the graph are due to true local fluctuations in diameter, as well as to spurious diameter measurements that resulted from planes perpendicular to the centerline path but oblique to the true colonic diameter. Areas where De = 0 (distance around path, 600 mm) were due to collapsed segments of the colon. (b) Graph depicts data for the MS method. Smoother representation of colonic distention is seen than with the CSA method owing to averaging over path distance. Areas of nondistention (*) are also depicted. (c) Graph depicts data with the SV method. Path is divided into five anatomic segments of the colon: rectum (R), sigmoid (S), descending (D), transverse (T), and ascending (A). De was averaged in each segment (dashed lines). Though with less spatial resolution than with other algorithms, general trends in distention are represented.

 

fig.ommitted  Figure 8b. Distention analysis in human colon. (a) Graph depicts data with the CSA method. Distention profile shows undulating baseline, which represents De from the rectum (point 0) to cecum (point 1520). High-frequency spikes (*) in the graph are due to true local fluctuations in diameter, as well as to spurious diameter measurements that resulted from planes perpendicular to the centerline path but oblique to the true colonic diameter. Areas where De = 0 (distance around path, 600 mm) were due to collapsed segments of the colon. (b) Graph depicts data for the MS method. Smoother representation of colonic distention is seen than with the CSA method owing to averaging over path distance. Areas of nondistention (*) are also depicted. (c) Graph depicts data with the SV method. Path is divided into five anatomic segments of the colon: rectum (R), sigmoid (S), descending (D), transverse (T), and ascending (A). De was averaged in each segment (dashed lines). Though with less spatial resolution than with other algorithms, general trends in distention are represented.

 

fig.ommitted  Figure 8c. Distention analysis in human colon. (a) Graph depicts data with the CSA method. Distention profile shows undulating baseline, which represents De from the rectum (point 0) to cecum (point 1520). High-frequency spikes (*) in the graph are due to true local fluctuations in diameter, as well as to spurious diameter measurements that resulted from planes perpendicular to the centerline path but oblique to the true colonic diameter. Areas where De = 0 (distance around path, 600 mm) were due to collapsed segments of the colon. (b) Graph depicts data for the MS method. Smoother representation of colonic distention is seen than with the CSA method owing to averaging over path distance. Areas of nondistention (*) are also depicted. (c) Graph depicts data with the SV method. Path is divided into five anatomic segments of the colon: rectum (R), sigmoid (S), descending (D), transverse (T), and ascending (A). De was averaged in each segment (dashed lines). Though with less spatial resolution than with other algorithms, general trends in distention are represented.

 

	Discussion

Top ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Gaseous distention of the colon is a critical determinant of diagnostic quality at CT colonography because the bowel is collapsed without distention, and polyps along the mucosal surface become indistinguishable from neighboring haustral folds and colonic wall (8,9,20). To optimize distention, investigators have proposed use of spasmolytics such as glucagon and butylscopolamine, and some have begun to use continuously insufflated rectal carbon dioxide as an alternative to room air (20). Before the additional time and expense of such techniques are accepted, validation of their effectiveness is needed. Moreover, such validation should ideally be on the basis of accurate, precise, and quantitative measurements of distention. It is possible that manual measurements of colonic diameter from cecum to rectum on transverse or reformatted images could provide the data necessary to distinguish among various distention protocols, but this would be extremely tedious and prone to operator error.

Our distention algorithms rely on initial processing of the CT data in the form of segmentation of air in the colon and determination of a centerline path. While each produces quantitative measures of D_e, the algorithms vary in their implementation, spatial resolution, accuracy, and sources of error. Their chief attributes are summarized in .

fig.ommitted TABLE 5. Summary of Attributes of Distention Measurement Algorithms

 
Probably the most intuitive method, the CSA method, produces accurate CSA measurements at each point along the centerline path and is highly sensitive to local changes in distention. In our colon models, even variations caused by surface features such as haustra were readily detected. As a result of this fine sampling, however, the CSA method produces relatively noisy distention profiles, which could make comparison between patients or patient positions difficult. In clinical CT colonography, one is mainly interested in distention on the basis of anatomic segments or subsegments as opposed to on a centimeter or millimeter spatial scale; therefore, this method likely provides an overly detailed distention profile. In the future, however, this type of high-spatial-resolution algorithm could be useful for semiautomatically detecting benign or malignant colonic strictures. Extension of the software to other areas, such as assessment of vascular stenoses or small bowel obstruction, may also be fruitful. Presently, the main weakness of the CSA method is the explicit dependence of the CSA planes on path orientation. Irregularities in the path or areas of high curvature may cause overestimation of the effective diameter. Such errors could be minimized if we implemented a subroutine in the algorithm to iteratively minimize the local CSA, as was done for the cut planes in the SV method. This would increase processing time.

The moving average of distention with the MS method algorithm creates a smoother distention profile than the CSA method, making trends in distention easier to grasp. Because of inherent averaging, however, the MS method is less accurate than the CSA method in representing rapid changes in diameter. The dampening effect can be limited by choosing the smallest possible sphere size that is larger than the colonic diameter. However, this is problematic because one would typically not know the colonic diameter before performing the algorithm. Another issue with the MS method is the length of time (several hours) that is required for computation. If rapid assessment is desired in clinical CT colonography, hours of processing time are not acceptable.

We believe that the SV algorithm is likely to be the most useful method for quantifying distention in the human colon. It is accurate in calculating mean segmental distention and by design gives the user flexibility in dividing the colon. While the SV method is not designed to describe local topology, it provides accurate depictions of regional and global distention. While a less sophisticated version of the SV algorithm could be useful for rapid assessment in clinical CT colonography, the end-plane optimization we implemented was critical in obtaining accurate volumes. Such a high level of accuracy is desirable for research trials that address the efficacy of various colon distention regimens.

We validated algorithms by using phantoms for several reasons. This approach made possible the comparison of experimental measurements with an explicit reference standard truth, which is not available from human CT colonographic data. Also, flexibility of the software simulator allowed efficient construction of a large number of models with a broad range of geometries.

In a blinded trial to assess the efficacy of glucagon, Yee et al (15) visually assessed distention in 33 patients treated with glucagon in comparison with 27 control subjects and found no appreciable effect on colonic distention. Nonetheless, whether or not there is a true benefit remains controversial, and some groups continue to routinely administer spasmolytics. Application of semiautomated quantitative methods should enable more rigorous testing of drug effects (21). Whereas room air is simply and inexpensively applied with a manual bulb insufflator, carbon dioxide requires a more expensive continuous insufflation device because it is absorbed more quickly by the colon (20). On the beneficial side, this rapid dissipation may make carbon dioxide better tolerated by the patient. Before investing in the necessary equipment for administering carbon dioxide, it would obviously be useful to document a beneficial effect, which could be accomplished by using our methods.

Quantitative methods can potentially be used as tools to assess how much of the colon surface is visible to the radiologist for evaluation in a given CT colonographic study. If analysis shows that a considerable portion of the colon is poorly distended, this increases the possibility that polyps could be overlooked. Since most investigators now scan fully with the patient in both supine and prone positions, it would be beneficial to combine the distention information from both positions. Currently, this is possible by superimposing the distention graphs from supine and prone positions. In the future, more sophisticated means of combining information from positions should be possible. Finally, distention analysis, if performed rapidly (such as with the CT scanner), could allow more selective scanning in the patient. While the literature supports scanning in both prone and supine positions (9,13,15), this doubles both the radiation dose and the number of images acquired. Rapid semiautomatic distention analysis could determine if a particular colon was adequately distended after one scan was acquired and help direct limited repeat scanning in another position.

Our study has several limitations. The path-planning and segmentation software used in the algorithms is a custom application that is not yet widely available. However, others have developed similar methods for creating centerlines and performing segmentation, and these could be similarly extended (22–24). Another issue is that the algorithms were designed to measure distention on the basis of the air-filled colon; thus, fluid in the test colon was interpreted as intestinal wall. Hence, our effective diameter determinations are an estimated rather than absolute diameter. While retained fluid is an important determinant of data quality at CT colonography (11), this separate detection and segmentation of fluid from the colon wall is very difficult because of the small CT density differences involved. One potential solution is to have patients ingest liquid iodinated or barium contrast material before the study (25,26). If residual fluid and feces are so labeled, computer-based removal of the high-density material is possible, and more accurate absolute diameters could be determined. Another limitation of this work is that our phantoms may not adequately mimic all features of the human colon. On the basis of our initial success in the application of the algorithms to a human CT colonographic data set, however, we are confident that clinical research applications will be practical (21).

In conclusion, we developed and measured in phantom models the accuracy of three computer algorithms for quantitative assessment of bowel distention in CT colonography. The choice of algorithm(s) for clinical or research applications will depend on how finely one desires to sample the distention profile and the absolute accuracy required. With systematic application to human data in clinical trials, we believe these algorithms can help optimize data acquisition at CT colonography.

APPENDIX
Herein, we provide a complete description of the computer algorithms.

CSA Method
Since the colonic lumen is approximately circular, D_e can be derived by using tomographic planes oriented perpendicular to the colonic axis or centerline. Implementation of this process requires segmentation of the colonic lumen, calculation of the centerline, computation of tomograms perpendicular to the centerline at each path point, and calculation of the CSA and cross-sectional diameter of the tomogram. Existing software was used for segmentation and centerline determination. At each path point, an oblique plane was then computed, and the intersection of the oblique plane with the segmented colonic volume defined the local CSA of the colon. To provide the most accurate measurement of CSA, the CSA method took into account partial volume averaging of voxels at the boundary zone between air and the bowel wall. For this, two user-defined threshold variables T₁ and T₂ were defined, with T₁ less than T₂. T₁ represents the level of attenuation in Hounsfield units below which a pixel is entirely included in the area measurement (-900 HU in the current experiment). T₂ represents the level of attenuation above which a pixel is entirely excluded from area measurement (set at -200 to -350 HU). Partial volume contributions to total area were linearly interpolated if the attenuation, T_p, of a given voxel was between T₁ and T₂ by using the following formula:

where A_C is contributed area and A_P is pixel area.

Finally, D_e was calculated from the CSA (A) as a function of the distance along the path by using the following formula:

These computations resulted in measurements of D_e at each point along the centerline path through each object.

MS Method
The motivation behind the MS method was to produce some smoothing of the graphic depictions of distention as a function of path distance relative to those produced with the CSA method. At the same time, we desired a method that would remain sensitive to local changes in distention by calculating a moving average of D_e.

This algorithm requires both a centerline path and a segmentation of the colon. The MS method generates a sphere centered at each path point and then computes the volume of intersection between the sphere and the local colonic segmentation (Fig 3). The colonic segment is modeled as a cylinder to allow calculation of its volume. From standard geometric relationships, the effective radius, r, of the colonic cylinder intersecting the sphere is given with the following formula:
here R is the radius of the sphere, and V is the colonic volume intersecting with the sphere, the latter being known from counting the voxels within the segmentation. With use of these computations, the MS method calculates D_e (2r) sequentially for a sphere centered at each path point, which results in a plot of D_e as a function of path distance. Note that unlike the CSA method, segmentation in the MS method was based on a discrete threshold of -700 HU as the boundary between air and the wall; thus, partial volume averaging at the boundary voxels was not explicitly accounted for. The rationale for this was to minimize computation time and to recognize the fact that this algorithm already provided averaging along the colonic path. Sphere size is an adjustable parameter set by the user before performing the MS method. To provide an accurate estimate of D_e, the sphere diameter must be larger than that of the colonic lumen. If the diameter of a sphere is smaller than that of the colon, the MS method may underestimate D_e.

SV Method
The SV method requires preprocessing in the form of segmentation and creation of a centerline path. From these, the algorithm divides the colon into separate subsegments, by splitting the segmentation along cut planes located at user-defined path points. To minimize errors that might occur owing to oblique cut planes at the ends of the colonic segments, a subroutine was implemented that iteratively adjusted the orientation of each cut plane until the minimum local colonic CSA was achieved for that plane. Subsequently, the algorithm calculated the colonic volume between each set of cut planes by counting the number of voxels. As with the MS method, segmentation with the SV method was based on a discrete threshold of -700 HU as the boundary between air and the wall. Since the number of path points between the cut planes was known, the SV could then be divided by the segmental path distance, which resulted in an average effective CSA. This effective area was then converted to an effective colonic diameter by using the formula for the area of a circle, as in the CSA method.

　

	REFERENCES

Top ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 

1. Greenlee RT, Murray T, Bolden S, Wingo PA. Cancer statistics, 2000. CA Cancer J Clin 2000; 50:7-33. [Abstract/Free Full Text]

2. Fenlon HM, Nunes DP, Schroy PC, III, Barish MA, Clarke PD, Ferrucci JT. A comparison of virtual and conventional colonoscopy for the detection of colorectal polyps. N Engl J Med 1999; 341:1496-1503[Erratum: N Engl J Med 2000; 342:524.].

3. Hara AK, Johnson CD, Reed JE, et al. Detection of colorectal polyps with CT colography: initial assessment of sensitivity and specificity. Radiology 1997; 205:59-65.

4. Yee J, Akerkar GA, Hung RK, Steinauer-Gebauer AM, Wall SD, McQuaid KR. Colorectal neoplasia: performance characteristics of CT colonography for detection in 300 patients. Radiology 2001; 219:685-692.

5. Dachman AH, Kuniyoshi JK, Boyle CM, et al. CT colonography with three-dimensional problem solving for detection of colonic polyps. AJR Am J Roentgenol 1998; 171:989-995.

6. Macari M, Milano A, Lavelle M, Berman P, Megibow AJ. Comparison of time-efficient CT colonography with two- and three- dimensional colonic evaluation for detecting colorectal polyps. AJR Am J Roentgenol 2000; 174:1543-1549.

7. Beaulieu CF, Jeffrey RB, Jr, Karadi C, Paik DS, Napel S. Display modes for CT colonography. II. Blinded comparison of axial CT and virtual endoscopic and panoramic endoscopic volume-rendered studies. Radiology 1999; 212:203-212.

8. Fenlon HM, Ferrucci JT. Virtual colonoscopy: what will the issues be? AJR Am J Roentgenol 1997; 169:453-458.

9. Fletcher JG, Johnson CD, Welch TJ, et al. Optimization of CT colonography technique: prospective trial in 180 patients. Radiology 2000; 216:704-711.

10. McFarland EG, Brink JA. Helical CT colonography (virtual colonoscopy): the challenge that exists between advancing technology and generalizability. AJR Am J Roentgenol 1999; 173:549-559.

11. Macari M, Lavelle M, Pedrosa I, et al. Effect of different bowel preparations on residual fluid at CT colonography. Radiology 2001; 218:274-277.

12. McFarland EG, Brink JA, Pilgram TK, et al. Spiral CT colonography: reader agreement and diagnostic performance with two- and three-dimensional image-display techniques. Radiology 2001; 218:375-383.

13. Chen SC, Lu DS, Hecht JR, Kadell BM. CT colonography: value of scanning in both the supine and prone positions. AJR Am J Roentgenol 1999; 172:595-599.

14. Goei R, Nix M, Kessels AH, Ten Tusscher MP. Use of antispasmodic drugs in double contrast barium enema examination: glucagon or Buscopan? Clin Radiol 1995; 50:553-557.

15. Yee J, Hung RK, Akerkar GA, Wall SD. The usefulness of glucagon hydrochloride for colonic distention in CT colonography. AJR Am J Roentgenol 1999; 173:169-172.

16. Yee J, Hung RK, Steinauer-Gebauer AM, Geetanjali A, Wall SD, McQuaid KR. Colonic distention and prospective evaluation of colorectal polyp detection with and without glucagon during CT colonography (abstr). Radiology 1999; 213(P):256.

17. Karadi C, Beaulieu CF, Jeffrey RB, Jr, Paik DS, Napel S. Display modes for CT colonography. I. Synthesis and insertion of polyps into patient CT data. Radiology 1999; 212:195-201.

18. Paik DS, Beaulieu CF, Jeffrey RB, Rubin GD, Napel S. Automated flight path planning for virtual endoscopy. Med Phys 1998; 25:629-637.

19. Fleischmann D, Rubin GD, Paik DS, et al. Stair-step artifacts with single versus multiple detector-row helical CT. Radiology 2000; 216:185-196.

20. Vining DJ. Virtual colonoscopy. Semin Ultrasound CT MR 1999; 20:56-60.

21. Yee J, Steinauer-Gebauer AM, Beaulieu CF, Paik DS, Hung PW, Napel S. Clinical application of a computer automated distention quantification method for CT colonography (abstr). Virtual Colonoscopy, Second International Symposium Boston, Mass: Boston University, 2000; 112.

22. Zhang Z, Wang G, Brown BP, McFarland EG, Haller J, Vannier MW. Fast algorithm for soft straightening of the colon. Acad Radiol 2000; 7:142-148.

23. Bitter I, Sato M, Bender MA, Kaufman A, Wan M, Wax MR. Automatic, accurate and robust colon centerline algorithm (abstr). Radiology 2000; 217(P):370.

24. McFarland EG, Wang G, Brink JA, Balfe DM, Heiken JP, Vannier MW. Spiral computed tomographic colonography: determination of the central axis and digital unraveling of the colon. Acad Radiol 1997; 4:367-373.

25. Zalis ME, Hahn PF, Arellano RS, Mueller PR. Improving patient tolerance of colon cancer screening: digital bowel subtraction in CT colonography (abstr). Radiology 2000; 217(P):370.

26. Callstrom MR, Johnson CD, Fletcher JG, Reed JE, Wilson RN, Alquist DA. A feasibility study of CT colonography of the unprepped colon (abstr). Radiology 2000; 217(P):371.