Automated image-based colon cleansing for laxative-free CT colonography computer-aided polyp detection Marius George Lingurarua) Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, 10 Center Drive, Bethesda, Maryland 20892 and Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Medical Center 111 Michigan Avenue Northwest, Washington District of Columbia 20010

Neil Panjwani Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, 10 Center Drive, Bethesda, Maryland 20892

Joel G. Fletcher Department of Radiology, Mayo Clinic, 200 First Street Southwest, Rochester, Minnesota 55905

Ronald M. Summers Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, 10 Center Drive, Bethesda, Maryland 20892

(Received 26 July 2011; revised 31 October 2011; accepted for publication 1 November 2011; published 23 November 2011) Purpose: To evaluate the performance of a computer-aided detection (CAD) system for detecting colonic polyps at noncathartic computed tomography colonography (CTC) in conjunction with an automated image-based colon cleansing algorithm. Methods: An automated colon cleansing algorithm was designed to detect and subtract taggedstool, accounting for heterogeneity and poor tagging, to be used in conjunction with a colon CAD system. The method is locally adaptive and combines intensity, shape, and texture analysis with probabilistic optimization. CTC data from cathartic-free bowel preparation were acquired for testing and training the parameters. Patients underwent various colonic preparations with barium or Gastroview in divided doses over 48 h before scanning. No laxatives were administered and no dietary modifications were required. Cases were selected from a polyp-enriched cohort and included scans 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. The CAD system was run comparatively with and without the stool subtraction algorithm. Results: The dataset comprised 38 CTC scans from prone and/or supine scans of 19 patients containing 44 polyps larger than 10 mm (22 unique polyps, if matched between prone and supine scans). The results are robust on fine details around folds, thin-stool linings on the colonic wall, near polyps and in large fluid/stool pools. The sensitivity of the CAD system is 70.5% per polyp at a rate of 5.75 false positives/scan without using the stool subtraction module. This detection improved significantly (p ¼ 0.009) after automated colon cleansing on cathartic-free data to 86.4% true positive rate at 5.75 false positives/scan. Conclusions: An automated image-based colon cleansing algorithm designed to overcome the challenges of the noncathartic colon significantly improves the sensitivity of colon CAD by approximately 15%. [DOI: 10.1118/1.3662918] Key words: CTC, colon cancer, laxative-free, cleansing, heterogeneous stool, polyp detection

I. INTRODUCTION With the advent of large screening programs for colon cancer, computed tomography colonography (CTC) has become a viable tool for the noninvasive detection of polyps and a complement or alternative to optical colonoscopy (OC).1–3 CTC is widely recognized as a very sensitive and specific modality to identify colonic polyps.4 While OC can provide better visual confirmation of colonic polyps, it has a few downsides compared to CTC, such as the inability to detect polyps hidden behind folds or stool, and the discomfort and risks of the colonoscope. In CTC, oral contrast agents containing iodine and 6633

Med. Phys. 38 (12), December 2011

barium are typically combined with a cathartic bowel preparation based on sodium phosphate to highlight the colonic fluid and increase the sensitivity of polyp detection.5 In addition, stool can also be enhanced with tagging agents with partial cathartic cleansing6 or even no use of laxative.7,8 Although studies have demonstrated that a large portion (72%) of the patients would prefer CTC instead of OC,9 the poor patient acceptance of the bowel preparation remains common to both optical and typical cathartic virtual colonoscopy.10 For better patient compliance, image-based colon cleansing performed on fecal-tagged CTC allows the laxative-free detection of colonic polyps. However, laxative-free CTC

0094-2405/2011/38(12)/6633/10/$30.00

6633

6634

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

introduces many challenges and imaging artifacts, such as poorly and heterogeneously tagged stool10 (tagged stool mixed with air or untagged stool), thin-stool sticking to the colonic walls, pseudo-enhancement of colonic tissue,11 and partial volume effects. These artifacts contribute to the change in shape of the colonic structures and increase the number of false positives (FP) in detecting colonic polyps. The automated and reliable removal of tagged feces remains challenging due to high variability in contrast intake. We propose an automated algorithm to subtract stool, as an alternative to using laxatives, prior to the computer-aided detection of colonic polyps. Computer-aided detection (CAD) systems must show sufficient versatility to produce robust analysis on a large variety of data. In the case of colonography, CAD systems have shown high sensitivity and specificity for the detection of colonic polyps.3,12–22,35 It has been shown that the use of second-read CAD significantly improves polyp detection.22 While CAD systems for colonography get increasing popularity and endorsement from the radiological community, they have not been designed to cope with the presence of stool, although labeling the stool with high contrast agents replaces the use of laxatives and reduces patient discomfort. In that direction, the subtraction of tagged colonic fluid after cathartic bowel preparation has been shown to benefit the detection and segmentation of polyps.23,24 The diagnosis using cathartic-free bowel preparation7 has been a subject of high interest for the large scale acceptance of CTC investigations. Protocols without use of laxative and dietary restrictions are under investigation in clinical centers.25–27 Recent studies showed that noncathartic-tagged CTC compares favorably with OC, with improved sensitivity when electronic stool subtraction is employed.28 There has been limited work on image-based colon cleansing and mainly for the analysis of CTC data acquired after full cathartic bowel preparation.10 In recent years, several studies have shown good performance on homogeneously tagged fluid or even stool. They include methods based on: region growing,29 eigenvalues of the Hessian matrix,30,31 polynomial fitting,32 material and material-transition models,33 expectation maximization,34,36 and fuzzy connectedness and level sets.37 Alternatively, adaptive density correction and mapping were successfully used in a CAD scheme to detect polyps in cases with reduced cathartic cleansing.38,39 Although the reported results are adequate for the removal of fully tagged stool, a major challenge remains the ability to subtract stool that is not homogeneously tagged, as seen in Fig. 1, without leaving residuals that increase the FP polyp detections during diagnosis. Similar techniques are employed by leading CTC commercial interpretation software and CAD products. Using fly through imaging of the reconstructed colon, these products allow mimicking OC. For this purpose, the tagged colonic fluid is removed prior to visualization, but heterogeneously tagged-stool hampers the detection of polyps. Fewer publications addressed the cleansing of heterogeneously tagged stool as it appears in cathartic-free bowel preparation. In Refs. 40 and 41, shape analysis, mosaic Medical Physics, Vol. 38, No. 12, December 2011

6634

FIG. 1. A typical example of cathartic-free tagged CT of the colon with (a) homogeneously well-tagged stool, (b) thin linings of high-attenuation contrast agent along the colonic walls and folds, and (c) a pool of heterogeneous stool with irregular tagging and dark air bubbles.

decomposition, and level sets contributed to the removal of heterogeneous stool, but the added value for polyp detection was not shown. The use of dual-energy CT machines to detect tagged materials is another promising approach.42 However, these types of data are sparse and uncommon in current clinical practice. Alternatively, in Ref. 43 level sets and minimum variance were employed to cleanse CTC data in a parametric and empirical method associated with the CAD of polyps. To our knowledge, the only other attempt to combine CAD and image-based colon cleansing from cathartic-free bowel prep was presented by Na¨ppi et al. in Ref. 44, based on their previous work.39 Their stool removal was based on pseudo-enhancement correction and Bayesian virtual tagging, and the evaluation was performed on lesions larger than 6 mm. Despite these recent advances, the added value of automated colon cleansing for the detection of colonic polyps from CTC data remains to be proven. We present an automated algorithm for the removal of stool with tagging variability from cases with various colonic preparations. The method is locally adaptive to adjust for inconsistent tagging and combines intensity, shape, and texture analysis with probabilistic optimization. The algorithm is evaluated on cleansing colonic polyps partially or fully covered by tagged stool. The detection of colonic polyps is significantly improved by removing FP candidates related to heterogeneous stool. To our knowledge, this paper presents the one of the first evaluations of a fully automated system for the laxative-free CTC computer-aided polyp detection. II. MATERIALS AND METHODS II.A. Data

Thirty-eight CTC data (19 patients with matched prone/ supine scans) from patients that underwent cathartic-free bowel preparation were acquired on LightSpeed Ultra and Pro:16 scanners (GE Healthcare). Patients underwent various colonic preparations with barium or Gastroview in divided doses over 48 h before scanning. No laxatives were administered and no dietary modifications required. Cases were

6635

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

6635

FIG. 2. Methodology flow chart for the automated image-based colon cleansing. After the segmentation of the colon area, an initial labeling of the image into air, tissue, and stool is performed in the single material classification based on image intensity and texture features. Then, the quadratic regression models the transitions between materials and shape analysis corrects for the pseudo-enhancement of colonic folds. The probabilities of materials are updated during expectation maximization before subtracting the stool from the image. The steps of the algorithm are presented in detail below.

selected from a polyp-enriched cohort and included scans 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. Small untagged stool particles and unlabeled colonic fluid were present in the image data. There were a total of 44 confirmed polyps (22 unique polyps) detected and marked by an experienced radiologist. All polyps selected in the dataset were 10 mm or larger; polyps smaller than 10 mm were not marked in the database and ignored for this study. The data were split in a random fashion into 18 training cases (9 patients/18 polyps) and 20 testing cases (10 patients/26 polyps). An outline of the stool subtraction method can be found in Fig. 2. II.B. Colon segmentation

The first step in our algorithm segments the colon by locating the colonic air and identifying the adjacent tissue and tagged materials. Colonic air is identified by thresholding at  700 HU and eliminating the components which border the image in the transverse plane and superior position. The stool tagged regions are identified by thresholding at 250 HU and removing the spine and ribs via region growing from high intensity seeds automatically placed in a superior position. After joining the colonic air and tagged materials, a morphological dilation is performed to include any surrounding tissue. A median filter is applied to reduce noise while preserving edges. II.C. Adaptive thresholding and single material classification

From the segmented colon, the algorithm utilizes image intensity, I, gradient magnitude, GM, and texture information to classify voxels into one material type: air, tissue, stool, or unclassified. In a previous approach,32 material classes were detected via hard thresholds of I and GM found empirically from the image data. However, there is a large variability in appearance throughout the image data, or even in the CT of one patient, due to poor and heterogeneous tagging. This makes hard thresholds unreliable and irreproducible. Our method assumes that the intensity of air is constant across the data, and similarly the intensity of the well-tagged stool. However, a number of adaptive thresholds learn the values of parameters from the image data. To account for local variability, we propose to compute thresholds in bounding boxes automatically defined around Medical Physics, Vol. 38, No. 12, December 2011

areas of tagged stool. Then, the thresholds are adaptively computed from the histogram analysis of each bounding box using the method in Ref. 45, which allows separating the materials automatically based on their respective frequencies. Morphological opening was used to separate distinct regions of high intensity in the colon. The opening was performed as a sequence of one binary erosion and one dilation, in order to remove weakly connected components. The radius of the structuring element was about 2 mm and approximated by 3 pixels in the axial plane (x and y dimensions) and 2 pixels in the z direction to compensate for the image anisotropy. Then, each left connected component was enveloped in a rectangular box, and the histogram of the box was analyzed resulting in the adaptive intensity threshold (Io) in Table I. A typical example of material distributions is shown in Fig. 3. A typical example of material distributions is shown in Fig. 3. The high heterogeneity in the appearance of tagged stool makes the intensity-based distinction of classes challenging. We propose to discriminate the parts of heterogeneity (high texture or lack of uniformity), which in the colon correspond to tagged stool, using a texture operator. In this context, image texture refers to statistical information about the spatial arrangement of intensities in a region of the image. To isolate heterogeneous stool with high texture, we locally compute Haralick’s correlation (T),46 which is a derivation of the 14 classic texture features proposed by Haralick and originally used to classify satellite imagery, and compare it to the maximum texture response, Tmax. T shows a strong response to both high texture and sharp contrast and is computed at every nonair voxel as follows, where g(i,j) is the element in cell i, j of a normalized gray level cooccurrence matrix (lt and rt are the mean and standard deviation of the row sums) TABLE I. Single material adaptive threshold values. Classes Air Tissue Stool Unclassified

Thresholds I   700 I  Io and I   250 and GM  300 and T < 0.75Tmax I  Io and GM < 0.8*I or I > 1000 Everything else

Note: I is the image intensity, GM is the gradient magnitude, Io is the adaptive intensity threshold, and T is the texture (Haralick’s correlation).

6636

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

FIG. 3. A typical distribution of material histograms from one CTC case with tagged stool. The distributions are normalized by their maximum value. The high peaks of the unclassified distribution likely represent the transitions between tissue-stool and tissue-air, which are affected by pseudoenhancement and partial volume effect. The transition between air-stool has a more ambiguous distribution partly overlapping with the tissue histogram.

P T¼

ði; jÞgði; jÞ  lt

i;j

rt

:

(1)

II.D. Quadratic regression

In order to address the material boundaries, a modified version of the quadratic regression (QR) algorithm published in Ref. 32 was applied. This allows comparing the distributions of I and GM to a set of parabolic curves defining the three two-material transitions: tissue-air, stool-tissue, and stool-air. Because the intensities for tissue and air are relatively constant in comparison to the variably tagged stool, the tissue-air parabola is unvarying. However, the stooltissue and stool-air parabolas are defined by a local stool maximum (Smax), which is computed across the image by minimizing the least-squares estimate of the stool-air parabola only (Fig. 4). We found that computing separate local stool maxima for stool-air and stool-tissue was counterintuitive and produced unpredictable results. We also utilized a morphological distance transform to classify any stool-related voxel into thin-stool if its maximum distance to

6636

the nearest nonstool voxel was less than or equal to 2.5 mm. After classification, the partial probabilities of air, tissue, and stool were computed for each voxel based on intensity and Smax. Note that the parameters of the quadratic regression are estimated adaptively in our application using the local bounding boxes. This allows our approach to account for the high variability in tagging and image appearance. The results of the quadratic regression offer the initial probabilities of the material transitions, which are updated as explained in the following sections. II.E. Hessian shape analysis

To further address pseudo-enhancement, we studied local concavity by computing the Hessian matrix at every voxel. The polarity of the eigenvalues (k1  k2  k3) of the Hessian matrix indicates local extrema, while particular combinations of the eigenvalue magnitudes can be used to discriminate various shapes. We followed the formulation set forth by Sato et al.47 to identify dark sheets in order to enhance submerged folds. The Hessian response, Sr, was computed as the maximum across scales to selectively enhance both thin and medium-sized folds 8   kt c > > > ; ks  kt  0 1  > > jks j > > <  (2) wðks ; kt Þ ¼ kt c jks j > 1a ; < kt < 0;  > > k j j > a s > > > : 0; else  jk1 jwðk1 ; k2 Þwðk1 ; k3 Þ; k1 > 0 Sr ¼ ; (3) 0; else where a ¼ 0.25 to restrict the response from negative eigenvalues and c ¼ 1 to sharpen the response within a limited range. We adapt the equations from Ref. 47 for the enhancement of folds (thin layers of tissue) surrounded by bright tagged stool. Additionally, we introduce corrections for the pseudoenhancement of submerged folds to update the partial probability of tissue in these regions. These corrections allow retrieving tissue that was misclassified by the quadratic regression, as explained below. Pseudo-enhanced tissue was classified by applying a median filter to the histogram of Sr and selecting voxels with responses higher than 30% of the maximum value. Because submerged folds represent local minima within stool, the partial tissue probability, pth, was recalculated by comparing the intensity at a voxel, I, to the maximum local intensity, HSmax, as follows:   I : (4) pth ¼ 1  HSmax II.F. Expectation maximization and subtraction

FIG. 4. Quadratic regression. The three parabolas in the picture are based on (Ref. 33) with additional modifications introduced from the adaptive local computation of Smax. Medical Physics, Vol. 38, No. 12, December 2011

After the partial probabilities were computed, a Gaussian mixture model expectation maximization (EM) algorithm48 was used. In addition to the image intensity distribution, we introduce neighborhood conditions to update the

6637

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

probabilities in the image data.43 The probability, ptþ1 j;i , of voxel i being in class j at time t þ 1 time step is 2  2  2 3 t t t t I  l p  k i j j;i j;i Wi 6 7 qffiffiffiffiffiffiffiffiffiffi exp4 2 ptþ1 5; j;i ¼ t t t t 2r 2q j j;i 2p r q j j;i

(5) where Ii is the intensity for voxel i, ljt, rjt, and Wit are the mean, variance, and weight for class j at time step t, ptj;i is the probability of class j for voxel i at time t, and lastly qtþ1 j;i 1 and ktþ1 are the mean and variance of the probability for j;i 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 five iterations. EM is used to update and integrate the partial probabilities that result from quadratic regression and shape analysis. While the quadratic regression initiates the probabilities using empirical curves, the Hessian analysis improves the probabilities around submerged folds. EM integrates these observations with neighborhood constraints in a Bayesian framework. Given the updated partial probabilities, a connected component was performed joining voxels with high probability of tissue under the knowledge that all tissue should be connected. The input intensity was then rescaled into the range [1000, Ioriginal] according to its probability of tissue, pt, value as follows:

Inew ¼ ptðIoriginal þ 1000Þ  1000:

(6)

Thus, tagged materials disconnected from the colonic wall were removed while identified stool with a low probability of tissue was set to the intensity of air. Gaussian smoothing of r ¼ 0.7 mm was performed at the edges of air to ease the transition of subtracted stool into tissue.

II.G. Colonic polyp CAD

For the automated detection of colonic polyps, the CAD system presented in Ref. 3 was used on the CTC database with and without colon cleansing. First, the colonic wall and lumen are segmented, and the tagged colonic fluid is removed. To identify polyps, the surface of the colon is analyzed and shape features are computed, and the candidates are segmented. Detections are classified using support vector machines to reduce the number of FP and a numeric score (a probability between 0 and 1) is generated for each polyp candidate to be a true or false detection. For more detail please refer to Ref. 42. The classifier was trained using the training data and the results are reported on the test set. When image-based cleansing was employed, both the training and test sets were from automatically cleansed cathartic-free data. The same was valid for the uncleansed results with training and testing performed on cathartic-free data without image-based cleansing. This way, we ensured that training and testing Medical Physics, Vol. 38, No. 12, December 2011

6637

images exhibit similar characteristics around the colonic lesions and false positives. In the current CAD system, a new tool was incorporated for scatter correction, which was not presented in Ref. 42 and was proven to benefit the detection of colonic polyps submerged in contrast agent.11 The scatter component, Is, of image I is estimated via a convolution with a symmetric scatter function SF Is ¼ I  SF; exp  SF ¼

(7) ðx  uÞ2 þ ðy  uÞ2 r2SF pr2SF

! ;

(8)

where u and rSF are the scale-related parameters of SF computed adaptively from the object scale in the image. This allows SF to adapt and be spatially variant according to the density and distribution of the contrast agent in I. u is the normalized difference between the computed object scale at the particular voxel and the maximum possible object scale. Thus, it is used in the scatter function to delineate how much to adjust the intensity. u is used in both the x and y directions because object scale aims to find the largest hyperball, which is the same size in the x and y dimensions.11 Free-response receiver operating characteristic (FROC) curves are generated using the numeric score computed by the CAD system. FROC analysis, the standard method for evaluating CAD performance, shows graphically the range of the sensitivity of CAD (true positives fraction) for detecting colonic polyps versus the specificity or FP rate (number of false positives per scan) for different values of the classifier output. As noted above, the parameter of the classifier in our study is a numeric score computed by the SVM incorporated in the CAD system. The numeric score represents the probability of a detection to be a true polyp, as determined by training the SVM on data with similar characteristics to the test set. This parameter allows tuning the CAD system to reach a compromise between a higher sensitivity and a greater number of false positives. FROC curves are generated from the CAD system with and without image-based colon cleansing. The significance between the two FROC curves is computed using the jackknife alternative free-response receiver operating characteristic (JAFROC) analysis.49,50 JAFROC is arguably one of the most reliable methods for the analysis of free-response CAD data in two or more modalities. The significance of intermodality differences is determined by jackknifing cases and analyzing the pseudo-value matrix with a mixed model analysis of variance.49,50 In this study, the significance (pvalue) was assessed using the software package provided.51 For typical clinical use, an operating point on the FROC curve with fixed sensitivity and false-positive rate is determined at a reasonable compromise between the sensitivity and specificity of the CAD system. To avoid the arbitrariness of the operating point, it was chosen in the part where the compared FROC curves become approximately flat. Thus,

6638

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

6638

the operating point does not reflect brisk changes in sensitivity as the specificity increases on either FROC curve. We report the sensitivity and specificity of the methods at this operating point in the FROC curves. III. RESULTS III.A. Automated colon cleansing

The intermediate results of the automated method for image-based colon cleansing can be found in Fig. 5. After creating a mask of the colon and its surrounding tissue, the parameters of the materials in the image data are computed locally [Fig. 5(c)]. Areas of high texture are identified to avoid labeling poorly tagged stool as colonic tissue [Fig. 5(d)]. Then, single materials and mixed material (transitions and partial volume effects) are identified [Fig. 5(f)]. Shape analysis highlighted submerged colonic folds in Fig. 6. The algorithm removed heterogeneous stool (Figs. 6–8), thin-stool linings (Fig. 6), and poorly tagged stool (Fig. 7), while preserving polyps (Fig. 8) and colonic tissue. Figures 9 and 10 present some examples of flythrough colonic images (virtual colonoscopy). The visualization software (V3D-COLON) performs the automated subtraction of tagged material but is hampered by the heterogeneous tagging of stool, resulting in highly variable fecal appearance. The residuals, left after the subtraction of tagged fluid and homogeneously tagged stool, are visible in the flythrough images [Figs. 9(a) and 10(a)] and can be mistaken for colonic lesions by the shape analyzers typically used in CAD. Using the proposed automated colon cleansing technique, the virtual colonoscopy images become clean, without altering the shape of the colonic surface, folds and lesions [Figs. 9(b) and 10(b)]. In both Figs. 9 and 10, the flythrough data are matched with their corresponding CTC images before and after stool removal. III.B. Colonic polyp CAD on cathartic-free data

The CAD system was run on CTC scans with and without the new technique for stool subtraction. A significant improvement was registered when the CAD was employed to detect polyps after the automated colon cleansing (p ¼ 0.009). The comparative FROC curves are presented in Fig. 11. As expected, the automatic detection of polyps is influenced by the presence of stool in the scans, as heterogeneous stool represents a prevailing source of FP. The sensitivity of colonic polyp detection increased after colon cleansing from 70.5% to 86.4% TP rate at 5.75 FP/scan.

FIG. 5. Results are illustrated at the intermediate steps of the colon cleansing method. The original CT (A) is processed after colon mask segmentation in (B), adaptive thresholding in (C), texture computation in (D), single material classification in (E), and identification of material transitions in (F). In (C), the blocks represent regions of locally computed intensity thresholds; brighter colors represent higher values in (C) and (D). (D) Illustrates the identification of areas of high texture. (E) Depicts the classification of air, tissue and stool separated by unclassified materials (white). (F) Illustrates the classification of air, tissue, stool, and tissue-air, stool-air and tissue-stool transitions.

The method adaptively adjusts to tagging conditions, variable interpatient data and local intrapatient variability. Moreover, the algorithm is robust to different bowel preparation techniques, as emphasized in the description of our dataset. The colon cleansing technique employed local adaptive analysis of the image combined with shape and texture analysis to identify the materials in the data: air, tissue, and stool. The local computation of parameters avoided hard thresholds that are sensitive to variable bowel preparations and

IV. DISCUSSIONS We presented a method to automatically subtract tagged stool from cathartic-free CT data to aid in the detection of colorectal polyps. The automated image-based colon cleansing algorithm was particularly designed to address the challenges of local variability in tagging, such as poor tagging or lack of enhancement, pseudo-enhancement of submerged colonic tissue, and the heterogeneity of stool enhancement. The results are robust on fine details around folds, thin-stool linings on the colonic wall, near polyps and in large fluid pools. Medical Physics, Vol. 38, No. 12, December 2011

FIG. 6. An example of submerged folds in heterogeneous stool; (a) is the original CTC image, (b) the results of the Hessian shape analysis with enhancement of submerged folds, and (c) is the final after colon cleansing. Note the variability in tagging in (a) with air bubbles enclosed by stool and the correct preservation of the colonic folds in (c).

6639

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

6639

FIG. 7. An example of automated image-based colon cleansing. The original CTC image in (a) corresponds to Fig. 5; the result of our stool removal algorithm is illustrated in (b).

local tagging. Texture identified areas of heterogeneous stool to avoid mislabeling untagged stool as tissue and bias the computation of local histograms. After quadratic regression, shape analysis recognized thin folds that were submerged by tagged material to prevent their erroneous removal due to pseudo-enhancement and partial volume effect. Finally, a maximum a posteriori probability scheme was employed with intensity and neighborhood terms to reduce the effect of noise and discontinuities in the regions of mixed materials. The scatter correction module in the CAD system generally corrected for areas of scatter and pseudo-enhancement in the images, but struggled with thin submerged folds and regions of enhanced heterogeneity in the tagged stool, where well-tagged, poorly tagged, and untagged stool mixed with air bubbles. We noted that on occasions, the scatter correction increased local heterogeneity in stool. To address these challenges, we used texture and shape analysis, as summarized above. Hessian shape analysis was used in the past in colonic image-based analysis30,41 where rut and cuplike structures were emphasized. We found no advantage in using our own implementation of the algorithm in Ref. 30 over a more traditional and intuitive identification of sheetlike structures that we adapted from Ref. 47. Additionally, our database

FIG. 8. Two examples of CTC data with polyps (circles) before (a and c) and after (b and d) automated colon cleansing. Medical Physics, Vol. 38, No. 12, December 2011

consisted of colonic lesions larger than 10 mm and often with irregular shapes that did not benefit from enhancing cuplike structures. Instead, we used texture analysis to help distinguish badly tagged stool from colonic tissue and maximum a posteriori probabilities to correct areas of transition between materials. To our knowledge, only three previous approaches have successfully removed heterogeneous stool from cathartic-free CTC data. The first one (Ref. 41) was discussed briefly above.

FIG. 9. A typical example of flythrough virtual colonoscopy before (A) and after (B) the automated colon-cleansing. Pictures on the bottom row show the location of the virtual colonoscope in the CTC image: (C) and (D) correspond to (A) and (B) respectively in the coronal view. Images were generated using the V3D-Colon visualization package [Viatronix, Stony Brook, NY] and approximated to match by the location and camera view. Note that (A) shows the flythrough after the automated removal of tagged material by the visualization software (V3D-Colon) with residuals, while (C) shows the uncleansed data. (B) and (D) present the automated cleansing results using the proposed method. The arrow indicates the position of the virtual camera.

6640

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

6640

FIG. 11. The comparative FROC curves using colonic polyp CAD on cathartic-free data with and without automated colon cleansing. The results are significantly improved when the proposed method for colon cleansing is employed.

FIG. 10. An example of flythrough virtual colonoscopy with an image of a colonic lesion. Pictures on the bottom row show the location of the virtual colonoscope in the CTC image: (C), (E) and (F) correspond to (A) and (D), (G) and (H) correspond to (B) in axial, coronal and sagittal views, respectively. Images were generated using the V3D-Colon visualization package [Viatronix, Stony Brook, NY] and approximated to match by the location and camera view. Note that (A) shows the flythrough after the automated removal of tagged material by the visualization software (V3D-Colon) with residuals, while (C), (E) and (F) show the uncleansed data. (B), (D), (G) and (H) present the automated cleansing results using the proposed method. The arrow indicates the position of the virtual camera. (E), (F), (G) and (H) are close-up views around the colonic lesion.

A second approach used dual-energy CT,42 an unusual type of CTC data in clinical practice. The third is our previous approach to remove heterogeneous stool43 that suffered from heavy parameterization and was based on minimum variance and curved level sets. With our current approach, we compute image parameters adaptively and involve shape and texture analysis to avoid common pitfalls in image-based colon cleansing. Importantly, the only attempt to automatically detect colonic polyps from cathartic-free CTC was presented in our previous publication,43 though it struggled with reproducibility. The CAD system detected colonic polyps with adequate accuracy on data with cathartic-free bowel preparation without employing the colon cleansing algorithm. However, the performance of the CAD system improved significantly after the cleansing module was applied. Employing the proposed technique in conjunction with CAD, the manuscript presents one of the first evaluations of a fully implemented automated system for the laxative-free CTC computer-aided polyp detection. The runtime for our cleansing algorithm averages 5 min R XeonV R processor (Intel, Santa per case on a Dual-Core IntelV Clara, CA) with 3.0 GHz and 8 GB RAM, while the CAD Medical Physics, Vol. 38, No. 12, December 2011

system approaches 15 min per case. We noticed, however, that the CAD system runtime decreased to roughly 10 min per case after colon cleansing due to the substantial decrease in false detections. The implementation used VISUAL Cþþ 8.0 (Microsoft, Redmond, WA) and ITK 3.20 (National Library of Medicine, Bethesda, MD). A current limitation of the method is the inability to cleanse correctly occasional areas of mixed tagging, as depicted in Fig. 12. In the case shown in Fig. 12, the untagged stool resembles submerged folds both in shape and appearance and is directly attached to the colonic wall. The size and shape of poorly tagged and/or untagged stool particles likely confuses the texture and probability analysis. Small holes can be created in thin folds after cleansing due to pseudoenhancement, but these artifacts do not resemble polyps in shape. Another reason for erroneous results may be the variable tagging procedures used in the data in our study, a shortcoming that is due to be fixed with the rapid advancement of tagging materials. In the future, we hope to reduce tagging variability by utilizing a single improved type of bowel preparation. The study lacked ground truth and analysis of polyps of 6–9 mm. Our images are from an anonymized dataset of patients acquired for a primary project focused on improving the laxative-free bowel preparation for detecting larger lesions. Further work will be required to show acceptable results for the clinically important 6–9 mm polyps, which was not possible in the current study. Instead, the results for

FIG. 12. An example of erroneous colon cleansing. (a) is a CTC image presenting poorly tagged heterogeneous stool and (b) shows the erroneous cleansing. The untagged stool resembles submerged folds both by shape and appearance and is directly attached to the colonic wall, which makes its removal very challenging.

6641

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

10 mm and larger polyps are comparable to systems in current clinical use.22 To summarize, the automated image-based colon cleansing technique significantly improved the computer-aided detection of colonic polyps from laxative-free CTC data. A system that allows the robust and accurate detection of colonic polyps from laxative-free CTC has the potential to increase the patient’s comfort and compliance with colon cancer screening protocols. ACKNOWLEDGMENTS This work was supported in part by the Intramural Research Program of the National Institutes of Health, Clinical Center and National Cancer Institute, Center for Cancer Research. The flythrough visualization was performed using V3D Colon 2.1, donated by Viatronix (Stony Brook, NY). a)

Author to whom correspondence should be addressed. Electronic mail: [email protected] 1 P. B. Cotton et al., “Computed tomographic colonography (virtual colonoscopy): A multicenter comparison with standard colonoscopy for detection of colorectal neoplasia,” Jama 291(14), 1713–1719 (2004). 2 H. B. Bosworth et al., “Prospective comparison of patient experience with colon imaging tests,” Am. J. Med. 119, 791–799 (2006). 3 R. M. Summers et al., “Computed tomographic virtual colonoscopy computer-aided polyp detection in a screening population,” Gastroenterology 129(6), 1832–1844 (2005). 4 C. D. Johnson et al., “Accuracy of CT colonography for detection of large adenomas and cancers,” N. Engl. J. Med. 359, 1207–1217 (2008). 5 P. J. Pickhardt et al., “Computed tomographic virtual colonoscopy to screen for colorectal neoplasia in asymptomatic adults,” N. Engl. J. Med. 349, 2191–2200 (2003). 6 P. Lefere, S. Gryspeerdt, J. Dewyspelaere, M. Baekelandt, and B. Van Holsbeeck, “Dietary fecal tagging as a cleansing method before CT colonography: Initial results—Polyp detection and patient acceptance,” Radiology 224(2), 393–403 (2002). 7 M. Callstrom, D. Johnson, J. Fletcher, J. Reed, D. Ahlquist, W. Harmsen, K. Tait, L. Wilson, and K. Corcoran, “CT colonography without cathartic preparation: Feasibility study,” Radiology 219(5), 693–698 (2001). 8 M. E. Zalis, J. Perumpillichira, and P. F. Hahn, “Digital subtraction bowel cleansing for CT colonography using morphological and linear filtration methods,” IEEE Trans. Med. Imaging 23(11), 1335–1343 (2004). 9 T. M. Gluecker et al., “Colorectal cancer screening with CT colonography, colonoscopy, and double-contrast barium enema examination: Prospective assessment of patient perceptions and preferences,” Radiology 227(2), 378–384 (2003). 10 R. M. Summers, “The elephant in the room: Bowel preparation for CT colonography,” Acad. Radiol. 16, 777–779 (2009). 11 J. Liu, J. Yao, and R. M. Summers, “Scale-based scatter correction for computer-aided polyp detection in CT colonography,” Med. Phys. 35(12), 5664–5671 (2008). 12 L. Bogoni et al., “Computer-aided detection (CAD) for CT colonography: A tool to address a growing need,” Br. J. Radiol. 78(1), S57–S62 (2005). 13 M. Fiori, P. Muse, and G. Sapiro, “A complete system for polyps flagging in virtual colonoscopy,” (2011) (IMA Preprint series 2367). 14 D. S. Paik et al., “Surface normal overlap: A computer-aided detection algorithm, with application to colonic polyps and lung nodules in helical CT,” IEEE Trans. Med. Imaging 23(6), 661–675 (2004). 15 V. F van Ravesteijn, C. van Wijk, R. Truyen, and J. Stoker, “Computer aided detection of polyps in CT colonography using logistic regression,” IEEE Trans. Med. Imaging 29(1), 120–131 (2009). 16 G. Slabaugh et al., “A robust and fast system for CTC computer-aided detection of colorectal lesions,” Algorithms 3(1), 21–43 (2010). 17 P. Sundaram, A. Zomorodian, C. Beaulieu, and S. Napel, “Colon polyp detection using smoothed shape operators: Preliminary results,” Med. Image Anal. 12(2), 99–119 (2008). 18 C. van Wijk, V. F. van Ravesteijn, F. M. Vos, and L. J. van Vliet, “Detection and segmentation of colonic polyps on implicit isosurfaces by Medical Physics, Vol. 38, No. 12, December 2011

6641

second principal curvature flow,” IEEE Trans. Med. Imaging 29(3), 688–698 (2010). 19 D. Heresbach et al., “Accuracy of computed tomographic colonography in a nationwide multicentre trial, and its relation to radiologist expertise,” Gut 60(5), 658–665 (2011). 20 J. W. Xu and K. Suzuki, “Massive-training support vector regression and Gaussian process for false-positive reduction in computer-aided detection of polyps in CT colonography,” Med. Phys. 38(4), 1888–1902 (2011). 21 H. Zhu, Y. Fan, H. Lu, and Z. Liang, “Improved curvature estimation for computer-aided detection of colonic polyps in CT colonography,” Acad. Radiol. 18(8), 1024–1034 (2011). 22 A. H. Dachman et al., “Effect of computer-aided detection for CT colonography in a multireader, multicase trial,” Radiology 256(3), 827–835 (2010). 23 R. M. Summers, M. Franaszek, M. T. Miller, P. J. Pickhardt, J. R. Choi, and W. R. Schindler, “Computer-aided detection of polyps on oral contrast-enhanced CT colonography,” Am. J. Roentgenol. 184(1), 105–108 (2005). 24 Y. Xu, W. Cai, J. Nappi, and H. Yoshida, “Fecal-tagging CT colonography with structure-analysis electronic cleansing for detection of colorectal flat lesions,” Eur. J. Radiol. (in press). 25 W. Davis, P. Nisbet, C. Hare, P. Cooke, and S. A. Taylor, “Non-laxative CT colonography with barium-based faecal tagging: Is additional phosphate enema beneficial and well tolerated?,” Br. J. Radiol. 84(998), 120–125 (2011). 26 S. Mahgerefteh, S. Fraifeld, A. Blachar, and J. Sosna, “CT colonography with decreased purgation: Balancing preparation, performance, and patient acceptance,” AJR, Am. J. Roentgenol. 193, 1531–1539 (2009). 27 S. A. Taylor, A. Slater, D. N. Burling, E. Tam, R. Greenhalgh, L. Gartner, J. Scarth, R. Pearce, P. Bassett, and S. Halligan, “CT colonography: Optimization, diagnostic performance and patient acceptability of reduced-laxative regimens using barium faecal tagging,” Eur. Radiol. 18, 32–42 (2008). 28 C. D. Johnson, A. Manduca, J. G. Fletcher, R. L. MacCarty, M. J. Carston, W. S. Harmsen, and J. N. Mandrekar, “Noncathartic CT colonography with stool tagging: Performance with and without electronic stool subtraction,” AJR, Am. J. Roentgenol. 190, 361–366 (2008). 29 D. Chen, Z. Liang, M. R. Wax, L. Li, B. Li, and A. E. Kaufman, “A novel approach to extract colon lumen from CT images for virtual colonoscopy,” IEEE Trans. Med. Imaging 19(12), 1220–1226 (2000). 30 W. Cai et al., “Structure-analysis method for electronic cleansing in cathartic and noncathartic CT colonography,” Med. Phys. 35(7), 3259–3277 (2008). 31 M. Oda et al., “Digital bowel cleansing free colonic polyp detection method for fecal tagging CT colonography,” Acad. Radiol. 16, 486–494 (2009). 32 M. Carston, A. Manduca, and C. D. Johnson, “Electronic stool subtraction using quadratic regression, morphological operations, and distance transforms,” Proc. SPIE. 6511(Part 1), 65110W (2007). 33 I. W. O. Serlie, “Electronic cleansing for computed tomography (CT) colonography using a scale-invariant three-material model,” IEEE Trans. Biomed. Eng. 57(6), 1306–1317 (2010). 34 Z. Wang et al., “An improved electronic colon cleansing method for detection of colonic polyps by virtual colonoscopy,” IEEE Trans. Biomed. Eng. 53(8), 1635–1646 (2006). 35 C. Robinson, S. Halligan, G. Iinuma, W. Topping, S. Punwani, L. Honeyfield, and S. A. Taylor, “CT colonography: Computer-assisted detection of colorectal cancer,” Br. J. Radiol. 84(1001), 435–440 (2011). 36 S. Wang, L. Li, H. Cohen, S. Mankes, J. J. Chen, and Z. Liang, “An EM approach to MAP solution of segmenting tissue mixture percentages with application to CT-based virtual colonoscopy,” Med. Phys. 35(12), 5787–5798 (2008). 37 M. Franaszek, R. M. Summers, P. J. Pickhardt, and J. R. Choi, “Hybrid segmentation of colon filled with air and opacified fluid for CT colonography,” IEEE Trans. Med. Imaging 25(3), 358–368 (2006). 38 J. Na¨ppi and H. Yoshida, “Fully automated three-dimensional detection of polyps in fecal-tagging CT colonography,” Acad. Radiol. 14, 287–300 (2007). 39 J. Na¨ppi and H. Yoshida, “Adaptive correction of CT attenuation for fecaltagging CT colonography,” Med. Image Anal. 12(4), 413–426 (2008). 40 W. Cai, H. Yoshida, M. E. Zalis, J. J. Na¨ppi, and G. J. Harris, “Electronic cleansing for non-cathartic CT colonography: A structure-analysis scheme,” Radiographics 30, 585–602 (2010). 41 W. Cai et al., “Mosaic decomposition: An electronic cleansing method for inhomogeneously tagged regions in noncathartic CT colonography,” IEEE Trans. Med. Imaging 30(3), 559–574 (2011). 42 R. Eliahou, Y. Azraq, R. Carmi, S. Y. Mahgerefteh, and J. Sosna, “Dualenergy based spectral electronic cleansing in non-cathartic computed

6642

Linguraru et al.: Automated image-based colon cleansing for laxative-free CT colonography

tomography colonography: An emerging novel technique,” Semin. Ultrasound CT MR 31, 309–314 (2010). 43 M. G. Linguraru et al., “Heterogeneous stool removal for laxative-free diagnosis of colon cancer—FROC study,” MICCAI Workshop on Virtual Colonoscopy (MICCAI Workshop, New York, 2008), pp. 85–90. 44 J. Na¨ppi, W. Cai, and H. Yoshida, “Computer-aided detection of colorectal lesions for cathartic-free CT colonography,” IEEE ISBI (IEEE ISBI Conference, Paris, France, 2009), pp. 911–914. 45 N. Otsu, “A threshold selection method from gray level histograms,” IEEE Trans. Syst. Man Cybern. 9, 62–66 (1979). 46 R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features for image classification, systems, man and cybernetics,” IEEE Trans. Syst. Man Cybern. 3(6), 610–621 (1973).

Medical Physics, Vol. 38, No. 12, December 2011

47

6642

Y. Sato et al., “Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images,” Med. Image Anal. 2(2), 143–168 (1998). 48 G. J. McLachlan and T. Krishnan, The EM Algorithm and Extensions, Wiley Series in Probability and Statistics, Applied Probability and Statistics (Wiley, New York, 1997). 49 D. P. Chakraborty and L. H. L. Winter, “Free-response methodology: Alternate analysis and a new observer-performance experiment,” Radiology 174, 873–881 (1990). 50 D. P. Chakraborty, “Analysis of location specific observer performance data: Validated extensions of the jackknife free-response (JAFROC) method,” Acad. Radiol. 13(10), 1187–1193 (2006). 51 www.devchakraborty.com.