Proc MICCAI 2008 Workshop: Computational and Visualization Challenges in the New Era of Virtual Colonoscopy
HeterogeneousStoolRemoval for Laxative-free CT Colonography Marius George Linguraru1, Robert L. Van Uitert1, Shan Zhao1, Jiamin Liu1 Joel G. Fletcher2, C. Daniel Johnson3, and Ronald M. Summers1, 1
Diagnostic Radiology Dept. Clinical Center, National Institutes of Health, Bethesda, MD 2 Department of Radiology, Mayo Clinic, Rochester, MN 3 Department of Radiology, Mayo Clinic, Scottsdale, AZ, USA
[email protected]
Abstract.Labeling the stool with high contrast agents for virtual colonoscopy replaces the use of laxatives and reduces the patient discomfort. However, this procedure introduces additional challenges for the diagnosis, such as poorly tagged stool, stool sticking to colonic walls, and heterogeneous stool (tagged stool mixed with air or untagged stool). Our study proposes a robust algorithm for heterogeneous stool removal that can be potentially used as a preprocessing step for computer aided diagnosis systems in colonic cancer detection. Colonoscopy data are automatically cleansed of residual stool to enhance the polyp appearance for improved diagnosis. The algorithm uses expectationmaximization, quadratic regression, level sets and minimum variance. Results show stool removal accuracy on polyps which are partially or fully covered by stool. The automatic detection of polyps using our CAD system on cathartic-free data improves significantly from 76% to 86% TP rate at 6.3 FP/scan.
1 Introduction In recent years, computed tomographic colonography (CTC) has become a viable and noninvasive procedure for the detection of polyps [1, 2]. Moreover, studies have shown that a large portion (72%) of the patients would prefer CTC to optical colonoscopy (OC) [3]. An additional factor of discomfort is the use of laxative for the preparation of the radiological exam and the diagnosis using cathartic-free bowel preparation is under investigation in clinical centers. However, this procedure introduces additional challenges for the diagnosis, such as poorly tagged stool, stool sticking to colonic walls, and heterogeneous stool (tagged stool mixed with air or untagged stool). While computer-aided diagnosis (CAD) systems for colonography get increasing popularity and endorsement from the radiological community, they are not designed to cope with the presence of stool. There has been little work on stool subtraction from cathartic-free colonic CT data and it includes methods based on: eigenvalues of the Hessian matrix [4], polynomial fitting [5], and expectation maximization [6]. Although the reported results are adequate for the removal of fully tagged stool, a major challenge remains the ability to subtract stool that is heterogeneously tagged without leaving residuals that increase the false positive (FP) polyp detections during diagnosis. Alternatively, adaptive
- 85-
Proc MICCAI 2008 Workshop: Computational and Visualization Challenges in the New Era of Virtual Colonoscopy
density correction and mapping were successfully used in a CAD scheme to detect polyps in cases with reduced cathartic cleansing [7]. Our study proposes a robust algorithm for heterogeneous stool removal to be employed as a preprocessing step for CAD systems in colon cancer detection. The method employs expectation maximization, polynomial regression, level-sets and minimum variance. The combined application of these algorithms is new, as is the use of variance followed by level-set to determine heterogeneity.
2 M aterialsandM ethods CTC data from cathartic-free bowel preparation were acquired at the Department of Radiology at the Mayo Clinic, Rochester MN: on LightSpeed Ultra (for testing) and LightSpeed Pro:16 scanners (for training), both from GE Healthcare. Cases underwent various colonic preparations with barium or Gastroview. Data were selected from a polyp-enriched cohort and included cases in which at least 90% of the solid stool was visually estimated to be tagged and each colonic segment was distended in either the prone or supine view. CTC data with cathartic bowel preparations were acquired at three centers: the National Naval Medical Center, Bethesda, MD;the Walter Reed Army Medical Center, Washington, DC;and the Naval Hospital in San Diego, CA;on LightSpeed or LightSpeed Ultra scanners (GE Healthcare). Patients underwent a 24-hour colonic preparation that consisted of oral administration of sodium phosphate, bisacodyl, barium, and diatrizoate meglumine and diatrizoate sodium [2]. An outline of the methodology can be found in Fig 1.a. The implementation used Visual C++ 8.0(Microsoft) and ITK 3.6. Stool-tagged CT Image
Ai r Ti ssue
Expectation Maximization
Uncl assi f i ed St ool
Polynomial Regression Fitting Heterogeneous Stool Removal -1200
-200
Cleansed CT Image
800
1800
Intensity
(a)
(b)
Fig.1.The flow chart of the proposed automatic stool subtraction algorithm is shown in (a); (b) presents the typical normalized distribution of the classes after expectation maximization.
2. 1 ClassificationofTissue,StoolandAir viaExpectationM aximization The first step in our subtraction method utilized expectation maximization (EM) [8] to adjust the variable intensities that might occur in different CT images. We divided the image into four classes: air, tissue, stool, and unclassified residuals. These classes were determined through a conservative adaptation of the values in [5].
- 86-
Proc MICCAI 2008 Workshop: Computational and Visualization Challenges in the New Era of Virtual Colonoscopy
The probability, Pjt,i 1 of voxel i being in class j at the t+1 time step is Pjt,i 1 )
Wi t 2. , tj ' * tj ,i
& ( I ( - tj ) 2 ( Pjt,i ( +tj , i ) 2 # , ( ' exp $( i ! 2 ' , tj 2 ' * tj ,i !"
%$where I i is the intensity for voxel i, - tj , , tj and Wi t are the mean, variance and weight for class j at time step t, Pjt,i is the probability of class j for for voxel i at time t, and lastly * tj,i and +tj,i are the mean and variance of the probability for class j on the first order neighbors around voxel i. We utilized the neighborhood values to reduce the effect of noise and discontinuities. The EM was run for 5 iterations;the voxels with small Ptissue were assigned to the unclassified category. The cutoffs were determined from the training data. We only reclassified tissue voxels, because air and well-tagged stool are easily distinguishable, while the boundaries and heterogenous stool were addressed at a later step. Fig. 1.b presents the normalized histogram of the intensity distributions of the four classes. 2. 2 BoundaryCorrectionviaQuadraticRegression During this second step, we followed the method by Cartson et al [5], with additional changes. The proposed detection of boundaries, or unclassified regions, is based on a quadratic regression of gradient magnitude versus intensity. This was done by comparing the distance deviation from a set of empirical curves found to describe uniquely each of the three transition boundaries (air-tissue, tissue-stool and air-stool). We did not utilize any morphological operators, as in [5]. In addition, instead of comparing the raw distance from the regressed curves, we biased the curve toward tissue-stool and air-tissue rather than stool-air, as to retain more tissue. To preserve folds covered by stool, we stipulated that a thin stool area that was originally classified as tissue-stool boundary should not be marked as stool 2. 3 RemovalofHeterogeneousStoolviaM inimum Variance andLevelSets The last step in our electronic cleansing alogrithm is a novel method to subtract heterogeneous stool, which can be separated into three stages: detect regions connected to tissue;detect boundaries via level sets;and correct for regions classified as tissue, but not connected to the colon wall, as regions of heterogeneous stool have a large body, but are normally separated from the colon wall by well-tagged stool. A chamfer distance map was used to determine the connectedness of a tissue voxel to the colon body and determine which voxels are classified as tissue. In addition, we computed the minimum 2D variance for each voxel across nine distinctive planes that can be used to cut a 5x5x5 neighboring volume to classify more tissue voxels. Next, we employed a curved level set [9] and an edge potential map that was constructed using 1 ( /+i1 +i 2 0 / max /+i1 +i 2 0 , where +i1 and +i 2 are the first and second i
largest eigenvalues of the Hessian matrix at voxel i. The level set algorithm was run for 50 iterations and from its propagation we determined which voxels that are attached to tissue can be classified as tissue-boundary voxels. We assumed that tissue voxels can propagate into boundaries, but boundaries cannot become tissue. This
- 87-
Proc MICCAI 2008 Workshop: Computational and Visualization Challenges in the New Era of Virtual Colonoscopy
condition connects areas on the colon folds that were previously erroneously labeled as stool, while isolating regions containing heterogeneous stool. Finally, we eliminated the heterogeneous stool voxels connected to the boundary. As before, we used the chamfer map and the connectivity to heterogeneous stool to classifiy stool voxels, thuis time tuned to propagate heterogeneous stool.
3 Results The electronic subtraction algorithm was tested on 44 CTC scans from prone and supine scans of 22 patients containing 43 polyps larger than 10mm (22 unique polyps, if matched between prone and supine scans). Eleven polyps were partially or fully covered by stool and 9of them were correctly cleansed, as exemplified in Fig. 2. The algorithm removed heterogeneous stool (Fig. 2.a and 2.b), stool linings (Fig 2.b), and weakly tagged stool (Fig. 2.c), while preserving the polyps and the colonic tissue. Two polyps (1 matched polyp) were not satisfactorily cleansed, as shown in Fig. 2.d.
Fig.2.Four polyps under tagged stool in uncleansed images (top row) and the results of our stool removal algorithm (bottom row). Arrows point at stool areas and ellipses at polyps.
The data without cathartic bowel preparation were employed to analyze the performance of the CAD system for colon cancer detection. The system was run on the 44 CTC scans with and without the new module of stool subtraction. The comparative free-response receiver operating characteristic (FROC) curves are presented in Fig. 3. CADSR refers to the CAD system with the incorporated stool removal (SR) module. Tests were performed on cathartic-free (CF) cases, cases with laxative preparation (LP) and the combination of the two (All). As expected, the automatic detection of polyps is hampered by the presence of stool in the scans, as heterogeneous stool represents a prevailing source of FP. However, the sensitivity of the CAD system for detecting polyps is 84% at 8.8 FP/scan without using the stool subtraction. The FROC curve improves significantly after stool subtraction on CF data from 76% to 86% true positive (TP) rate at 6.3 FP/scan (p-value<0.05).
- 88 -
Proc MICCAI 2008 Workshop: Computational and Visualization Challenges in the New Era of Virtual Colonoscopy
Cases with cathartic bowel preparation from 162 CTC scans (81 patients with prone and supine scans) were used for comparative FROC analysis on combinations of data. There were a total of 49 polyps larger than 10mm. The best results are achieved by CAD on LP cases with 88% TP rate at 4.5 FP/scan. However, CADSR detects polyps with 84% sensitivity at 4.5 FP/scan on CF data. CAD and CADSR show very similar results on a combination of CF and LP cases with 70% sensitivity at 8.8 FP/scan. Not surprisingly, CADSR fails to detect accurately polyps from LP data with 55% TP rate at 9.5 FP/scan, as data without stool does not comply with the class distributions of air, tissue and stool. 1 0. 9 0. 8
Sensi t i vi t y
0. 7 0. 6 0. 5 0. 4
CAD_CF CAD_LP
0. 3
CAD_Al l 0. 2
CADSR_CF CADSR_LP
0. 1
CADSR_Al l 0 0
2
4
6
8
10
FP/ scan
Fig.3.FROC curves of the performance of the CAD system on combinations of CTC data; we present comparative performances with and without the use of the stool subtraction module on cathartic-free (CF) data, cases with laxative preparation (LP) and the combination of the two.
4 Discussion We presented a method to remove tagged stool from cathartic-free bowel preparation CT data to assist with the diagnosis of colonic cancer. The method was particularly designed to remove well-tagged, heterogenous, as well as weakly tagged stool. The results are robust on fine grain details around folds, stool lining on polyps and large pools of stool. The method allows for variability in the input data set. The electronic cleansing algorithm failed to cleanse a polyp (both in the supine and prone scans) due to the assumption that large air pockets do not exist in stool fragments. Future work will increase the robustness of our algorithm to accommodate more variability in the cases of heterogeneously-tagged stool. The CAD system [2] detected colonic cancer with fine accuracy on data with cathartic-free bowel preparation, without employing the new algorithm for
- 89-
Proc MICCAI 2008 Workshop: Computational and Visualization Challenges in the New Era of Virtual Colonoscopy
heterogeneous stool removal. However, the performance of the CAD system showed a significant improvement on stool-tagged data when the stool removal module was used, with results appropriate for routine clinical use. When tested on a combination of cases with and without cathartic bowel preparation, the CAD system showed similar results with and without electronic cleansing. The runtime for our digital stool subtraction varies between one and two hours per patient, depending on the data size and case complexity for convergence. However, we noted that the large number of FP generated without stool subtraction can cause the classifier to have a running time of two hours, and on average, the amount of time required to run a case with or without stool subtraction is approximately the same. Results indicate a strong incentive toward the construction of a dual CAD system that takes into account the type of bowel preparation. With a CAD trained on cathartic bowel preparation data and an additional module of electronic cleansing trained on cathartic-free bowel preparation data, the new system can detect colonic polyps accurately with sensitivity and specificity appropriate for routine clinical use. Acknowledgments. This work was supported in part by the Intramural Research Program of the National Institutes of Health, Clinical Center.
References 1. Cotton, P.B., et al., Computed tomographic colonography (virtual colonoscopy): a multicenter comparison with standard colonoscopy for detection of colorectal neoplasia. Jama, 2004. 291(14): 1713-9. 2. Summers, R.M., et al., Computed tomographic virtual colonoscopy computer-aided polyp detection in a screening population. Gastroenterology, 2005. 129(6):1832-44. 3. Gluecker, T.M., et al., Colorectal cancer screening with CT colonography, colonoscopy, and double-contrast barium enema examination: prospective assessment of patient perceptions and preferences. Radiology, 2003.227(2):378-84. 4. Cai W, et al. Structure-analysis method for electronic cleansing in cathartic and noncathartic CT colonography. Medical Physics 2008;35:3259-77 5. Carston, M., A. Manduca, and C.D. Johnson. Electronic stool subtraction using quadratic regression, morphological operations, and distance transforms. In Proceedings of SPIE. 2007, 6511(Part 1): 65110W. 6. Wang, Z., et al., An improved electronic colon cleansing method for detection of colonic polyps by virtual colonoscopy. IEEE Transactions on Biomedical Engineering, 2006. 53(8): 1635-1646. 7. Nappi, J. and H. Yoshida, Fully Automated Three-Dimensional Detection of Polyps in Fecal-Tagging CT Colonography. Acad Radiol, 2007. 14(3): 287-300. 8. McLachlan, G.J. and T. Krishnan, The EM algorithm and extensions. Wiley series in probability and statistics. Applied probability and statistics. 1997, New York: Wiley. 9. Lorigo, L.M., et al., CURVES: Curve evolution for vessel segmentation. Medical Image Analysis, 2001. 5(3): 195-206.
- 90 -