Computer-Aided Polyp Detection for Laxative-Free CT Colonography Neil Panjwani1, Marius George Linguraru1, Joel G. Fletcher2, and Ronald M. Summers1 1

Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, Bethesda, MD 2 Department of Radiology, Mayo Clinic, Rochester, MN {paniwani,mglinguraru}@gmail.com

Abstract. Image-based colon cleansing performed on fecal-tagged CT colonography (CTC) allows the laxative-free detection of colon polyps, unlike optical colonoscopy (OC), the preferred screening method. Compared to OC, CTC increases the patient comfort and compliance with colon cancer screening. However, laxative-free CTC introduces many challenges and imaging artifacts, such as poorly and heterogeneously tagged stool, thin stool close to the colon walls, pseudoenhancement of colon tissue, and partial volume effect. We propose an automated algorithm to subtract stool prior to the computer aided detection of colonic polyps. The method is locally adaptive and combines intensity, shape and texture analysis with probabilistic optimization. Results show stool removal accuracy on data with various bowel preparations. The automatic detection of polyps using our CAD system on cathartic-free data improves significantly from 70% to 85% true positive rate at 5.75 false positives/scan. Keywords: CTC, colon cancer, laxative-free, cleansing, heterogeneous stool, polyp detection.

1

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) [1,2]. While OC can provide better visual confirmation of colon polyps, it has a few downsides compared to CTC, such as the inability to detect polyps hidden behind folds, stool, or fluid and the discomfort and risks of the colonoscope. In addition, studies have demonstrated that a large portion (72%) of the patients would prefer CTC instead of OC [3]. Furthermore, in CTC stool can be enhanced with tagging agents such as barium or iodine, as opposed to the use of a laxative. The automated and reliable removal of tagged feces remains challenging due to high variability in contrast intake that leads H. Yoshida et al. (Eds.): Abdominal Imaging 2011, LNCS 7029, pp. 18–26, 2012. © Springer-Verlag Berlin Heidelberg 2012

Computer-Aided Polyp Detection for Laxative-Free CT Colonography

19

to poorly and heterogeneously tagged stool, thin stool close to the colon walls, pseudoenhancement of colon tissue, and partial volume effect. In recent years, several studies have been developed for image-based colon cleansing, which perform well on fully homogeneously tagged stool [4,5,8]. Fewer publications addressed the cleansing of heterogeneously tagged stool. In [9], shape analysis and mosaics contributed to the removal of heterogeneous stool, but the added value for polyp detection was not shown. Alternatively, in [7] level sets and minimum variance were employed to cleanse CTC data in a parametric and empirical method associated with the computer aided detection (CAD) of polyps. 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 detection of colon polyps is significantly improved by removing false positive candidates related to heterogeneous stool.

2

Materials and Methods

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). Cases underwent various colonic preparations of barium or Gastroview. The dataset was selected from an original set of 50 to include only those cases high in polyps, 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. There were a total of 44 confirmed polyps (22 unique polyps) detected by an experienced radiologist. All polyps selected in the dataset were 10mm or larger; polyps smaller than 10mm were ignored for this experiment. The data were split into 18 training cases (9 patients/ 18 polyps) and 20 testing cases (10 patients/ 26 polyps). An outline of the method can be found in Fig 1.

Colon Segmentation

Single Material Classification

Quadratic Regression

Hessian Shape Analysis

Expectation Maximization

Subtraction

Fig. 1. Methodology Flow Chart

2.1

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

20

N. Panjwani et al.

thresholding at -600 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. 2.2

Adaptive Thresholding and Single Material Classification

We aim to conservatively classify each voxel into one material type (air, tissue, stool, or unclassified) by utilizing image intensity, I, gradient magnitude, GM, and texture, T. In [4], material classes were detected via hard thresholds. The adaptive intensity threshold (Io) in Table 1 is computed by applying Otsu's method [10] to the histograms of each local stool region. Regions are selected by locating each set of connected voxels with intensity higher than 200 HU, performing a morphological opening to remove weak, thin connections, and including the rectangular bounding box of each set, with a padding of 5 voxels. These regions allow a unique Io for each large, distinct stool region while providing significant information to the adaptive histogram computation. To isolate heterogeneous stool via texture, we locally compute Haralick's Correlation [12], which is a derivation of the 14 classic texture features proposed by Haralick (originally used to classify satellite imagery). Because T shows a strong response to heterogeneity and sharp intensity contrast, higher measures of T are indicative of stool. We compare the value of T computed at a particular voxel to the maximum global texture response, Tmax, to normalize the texture measure. T is computed at every non-air voxel as follows, where , is the element in cell i, j of a normalized grey level co-occurrence matrix ( and are the mean and standard deviation of the row sums): ∑,

,

,

(1)

Table 1. Single material adaptive threshold values Classes Air Tissue Stool Unclassified

2.3

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

Quadratic Regression

To address material boundaries, a slightly modified version of the quadratic regression (QR) algorithm published in [4] was applied. Unclassified voxels are assigned a material transition by finding the shortest orthogonal distance from a set of points in the I vs. GM graph to the parabolic curves modeling tissue-air, stool-tissue,

Computer-Aided Polyp Detection for Laxative-Free CT Colonography

21

and stool-air transitions. 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 stool-tissue and stool-air parabolas are defined by a local stool maximum (Smax), which is computed across the image by minimizing the leastsquares estimate of the stool-air parabola only. We found that computing separate local stool maxima for stool-air and stool-tissue was counter-intuitive 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 the nearest non-stool voxel was less than or equal to 2.5mm. After classification, the partial probabilities of air, tissue, and stool were computed for each voxel based on intensity and Smax. 2.4

Hessian Shape Analysis

To address the pseudoenhancement of colonic folds, we studied local concavity by computing the Hessian matrix at every voxel. The polarity of the eigenvalues 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 [11] to identify dark sheets in order to enhance submerged folds surrounded by brightly tagged materials. The Hessian response, Sσ was computed as the maximum across scales to selectively enhance both thin and mediumsized folds: 1 ,

1

|

|

|

|

, ,

0 |

|

0

0, | |

,

,

, 0,

; (2)

0 (3)

where 0.25 to restrict the response from negative eigenvalues and 1 to sharpen the response within a limited range. Pseudoenhanced tissue was classified by applying a median filter to the histogram of Sσ 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: 1

2.5

(4)

Expectation Maximization and Subtraction

After the partial probabilities were computed, a Gaussian mixture model expectation maximization algorithm was used to update the probabilities on the basis of intensity

22

N. Panjwani et al.

and neighboring probabilities. The probability, time t+1 time step is:

,

, of voxel i being in class j at

, ,

2

2

,

2

,

,

,

(5)

where is the intensity for voxel i, , , and are the mean, variance and weight for class j at time step t, , is the probability of class j for voxel i at time t, and lastly , and , 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. 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, ] according to its probability of tissue, pt, value as follows:

1000

1000

(6)

Thus, tagged materials disconnected from the colon wall were removed while identified stool with a low probability of tissue were set to the intensity of air. Gaussian smoothing of σ = 0.7 mm was performed at the air edges to ease the transition of subtracted stool into tissue. 2.6

Colonic Polyp CAD

For the automated deception of colonic polyps, the CAD tool presented in [2] was used on the CTC database with and without colon cleansing. Results are compared using free-response operator characteristic (FROC) analysis.

3

Results

The intermediate results of the automated method for stool subtraction can be found in Fig. 2. The algorithm removed heterogeneous stool (Fig. 3,4,5), thin stool linings (Fig. 3), and weakly tagged stool (Fig. 4), while preserving polyps (Fig. 5) and colonic tissue.

Computer-Aid ded Polyp Detection for Laxative-Free CT Colonography

23

Fig. 2. The original CT (A) processed p after colon segmentation (B), adaptive thresholding (C), texture computation (D), singlee material classification (E), and quadratic regression (F). In (C), the blocks represent regions of locaally computed intensity thresholds. Brighter colors represent higgher values in figures (C) and (D). Fiigure (E) depicts the classification of air (blue), tissue (orange), sstool (green), and unclassified (whitee). Figure (F) illustrates the classification of air (dark blue), tiissue (light blue), stool (green), tissuee-air (orange), stool-air (yellow), tissue-stool (purple), and thin sstool (white).

Fig. 3. (A) Original CT. (B) After A Hessian shape analysis enhancement of submerged folds. (C) Final subtracted result.

24

N. Panjwani et al.

Fig. 4. The original CT (A) from Figure 2A and the result of our stool removal algorithm (B)

Fig. 5. CT images with polyps (green circles) before (A and C) and after (B and D) cleansing. The CAD system was run on the CTC scans with and without the new module for stool subtraction. The comparative FROC curves are presented in Fig. 7. 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 false positives (FP). The FROC curve improves significantly after stool subtraction from 70% to 85% true positive (TP) rate at 5.75 FP/scan (p = 0.007). The statistical significance was computed using ROCKIT.

Computer-Aided Polyp Detection for Laxative-Free CT Colonography

25

1 0.9 0.8

Sensitivity

0.7 0.6 0.5

Cleansed

0.4

Uncleansed

0.3 0.2 0.1 0 0

2

4

6

8

10

12

14

16

18

20

FP/scan Fig. 6. FROC curves of the performance of the CAD system before and after cleansing

Fig. 7. (A) CT image depicting poorly tagged heterogeneous stool. (B) Erroneous subtraction.

4

Discussion

We presented a method to automatically subtract tagged stool from cathartic-free CT data to aid in the detection of colorectal cancer. The stool subtraction algorithm was particularly designed to address the challenges of local variability in tagging, pseudoenhancement of submerged colon 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. The method adaptively adjusts to variable patient data and tagging conditions. However, the method relies on adequate stool tagging to work properly, as depicted in Fig. 7. The runtime for our cleansing algorithm averages 5 min per case, while the CAD system approaches 15 min per case. We noticed, however, that the CAD system runtime decreased to roughly 10 min per case after cleansing due to the elimination of several false detections. The CAD system detected colonic polyps with sufficient accuracy on data with cathartic-free bowel preparation without employing the cleansing algorithm. However, the performance of the CAD system improved significantly after the cleansing module was applied, with results appropriate for routine clinical use.

26

N. Panjwani et al.

Future work will address specific artifacts brought on by air bubbles within the stool and weak or no stool tagging. Cases with polyps under 10mm should be examined. We also hope to reduce tagging variability by utilizing a single improved type of bowel preparation. 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 291(14), 1713–1719 (2004) 2. Summers, R.M., et al.: Computed Tomographic Virtual Colonoscopy Computer-Aided Polyp Detection in a Screening Population. Gastroenterology 129(6), 1832–1844 (2005) 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 227(2), 378–384 (2003) 4. Carston, M., Manduca, A., Johnson, C.D.: Electronic Stool Subtraction Using Quadratic Regression, Morphological Operations, and Distance Transforms. In: Proceedings of SPIE 6511 (Part 1), 65110W (2007) 5. Wang, Z., 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) 6. McLachlan, G.J., Krishnan, T.: The EM Algorithm and Extensions. Wiley Series in Probability and Statistics. Applied Probability and Statistics. Wiley, New York (1997) 7. Linguraru, M.G., et al.: Heterogeneous Stool Removal for Laxative-Free Diagnosis of Colon Cancer – FROC Study. In: MICCAI Workshop on Virtual Colonoscopy, pp. 85–90 (2008) 8. Cai, W., et al.: Structure-Analysis Method for Electronic Cleansing in Cathartic and Noncathartic CT Colonography. Med. Phys. 35(7), 3259–3277 (2008) 9. Cai, W., 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) 10. Otsu, N.: A Threshold Selection Method from Gray Level Histograms. IEEE Trans. Syst. Man Cybern. 9, 62–66 (1979) 11. Sato, Y., et al.: 3D Multi-Scale Line Filter for Segmentation and Visualization of Curvilinear Structures in Medical Images. In: Troccaz, J., Mösges, R., Grimson, W.E.L. (eds.) CVRMed-MRCAS 1997, CVRMed 1997, and MRCAS 1997. LNCS, vol. 1205, pp. 213–222. Springer, Heidelberg (1997) 12. Haralick, R.M., Shanmugam, K.: Dinstein, Its’Hak: Textural Features for Image Classification. IEEE Trans. Syst. Man Cybern. 3(6), 610–621 (1973)