Computer-aided detection of polyps in CT colonography using logistic regression

We present a computer-aided detection (CAD) system for computed tomography colonography that orders the polyps according to clinical relevance. The CAD system consists of two steps: candidate detection and supervised classification. The characteristics of the detection step lead to specific choices for the classification system. The candidates are ordered by a linear logistic classifier (logistic regression) based on only three features: the protrusion of the colon wall, the mean internal intensity, and a feature to discard detections on the rectal enema tube. This classifier can cope with a small number of polyps available for training, a large imbalance between polyps and non-polyp candidates, a truncated feature space, unbalanced and unknown misclassification costs, and an exponential distribution with respect to candidate size in feature space. Our CAD system was evaluated with data sets from four different medical centers. For polyps larger than or equal to 6 mm we achieved sensitivities of respectively 95%, 85%, 85%, and 100% with 5, 4, 5, and 6 false positives per scan over 86, 48, 141, and 32 patients. A cross-center evaluation in which the system is trained and tested with data from different sources showed that the trained CAD system generalizes to data from different medical centers and with different patient preparations. This is essential to application in large-scale screening for colorectal polyps.

Polyp measurement based on CT colonography and colonoscopy: variability and systematic differences

Objective

To assess the variability and systematic differences in polyp measurements on optical colonoscopy and CT colonography.

Materials

Gastroenterologists measured 51 polyps by visual estimation, forceps comparison and linear probe. CT colonography observers randomly assessed polyp size two-dimensionally (abdominal and intermediate window) and three-dimensionally (manually and semi-automatically). Linear mixed models were used to assess the variability and systematic differences between CT colonography and optical colonoscopy techniques.

Results

The variability of forceps and linear probe measurements was comparable and both showed less variability than measurement by visual assessment. Measurements by linear probe were 0.7 mm smaller than measurements by visual assessment or by forceps. The variability of all CT colonography techniques was lower than for measurements by forceps or visual assessment and sometimes lower (only 2D intermediate window and manual 3D) compared with measurements by linear probe. All CT colonography measurements judged polyps to be larger than optical colonoscopy, with differences ranging from 0.7 to 2.3 mm.

Conclusion

A linear probe does not reduce the measurement variability of endoscopists compared with the forceps. Measurement differences between observers on CT colonography were usually smaller than at optical colonoscopy. Polyps appeared larger when using various CT colonography techniques than when measured during optical colonoscopy.

Automated detection and segmentation of large lesions in CT colonography

Computerized tomographic colonography is a minimally invasive technique for the detection of colorectal polyps and carcinoma. Computer-aided diagnosis (CAD) schemes are designed to help radiologists locating colorectal lesions in an efficient and accurate manner. Large lesions are often initially detected as multiple small objects, due to which such lesions may be missed or misclassified by CAD systems. We propose a novel method for automated detection and segmentation of all large lesions, i.e., large polyps as well as carcinoma. Our detection algorithm is incorporated in a classical CAD system. Candidate detection comprises preselection based on a local measure for protrusion and clustering based on geodesic distance. The generated clusters are further segmented and analyzed. The segmentation algorithm is a thresholding operation in which the threshold is adaptively selected. The segmentation provides a size measurement that is used to compute the likelihood of a cluster to be a large lesion. The large lesion detection algorithm was evaluated on data from 35 patients having 41 large lesions (19 of which malignant) confirmed by optical colonoscopy. At five false positive (FP) per scan, the classical system achieved a sensitivity of 78%, while the system augmented with the large lesion detector achieved 83% sensitivity. For malignant lesions, the performance at five FP/scan was increased from 79% to 95%. The good results on malignant lesions demonstrate that the proposed algorithm may provide relevant additional information for the clinical decision process.

Primary uncleansed 2D versus primary electronically cleansed 3D in limited bowel preparation CT-colonography. Is there a difference for novices and experienced readers?

The purpose of this study was to compare a primary uncleansed 2D and a primary electronically cleansed 3D reading strategy in CTC in limited prepped patients. Seventy-two patients received a low-fibre diet with oral iodine before CT-colonography. Six novices and two experienced observers reviewed both cleansed and uncleansed examinations in randomized order. Mean per-polyp sensitivity was compared between the methods by using generalized estimating equations. Mean per-patient sensitivity, and specificity were compared using the McNemar test. Results were stratified for experience (experienced observers versus novice observers). Mean per-polyp sensitivity for polyps 6 mm or larger was significantly higher for novices using cleansed 3D (65%; 95%CI 57–73%) compared with uncleansed 2D (51%; 95%CI 44–59%). For experienced observers there was no significant difference. Mean per-patient sensitivity for polyps 6 mm or larger was significantly higher for novices as well: respectively 75% (95%CI 70–80%) versus 64% (95%CI 59–70%). For experienced observers there was no statistically significant difference. Specificity for both novices and experienced observers was not significantly different. For novices primary electronically cleansed 3D is better for polyp detection than primary uncleansed 2D.

Does a computer-aided detection algorithm in a second read paradigm enhance the performance of experienced computed tomography colonography readers in a population of increased risk?

We prospectively determined whether computer-aided detection (CAD) could improve the performance characteristics of computed tomography colonography (CTC) in a population of increased risk for colorectal cancer. Therefore, we included 170 consecutive patients that underwent both CTC and colonoscopy. All findings >or=6 mm were evaluated at colonoscopy by segmental unblinding. We determined per-patient sensitivity and specificity for polyps >or=6 mm and >or=10 mm without and with computer-aided detection (CAD). The McNemar test was used for comparison the results without and with CAD. Unblinded colonoscopy detected 50 patients with lesions >or=6 mm and 25 patients with lesions >or=10 mm. Sensitivity of CTC without CAD for these size categories was 80% (40/50, 95% CI: 69-81%) and 64% (16/25, 95% CI: 45-83%), respectively. CTC with CAD detected one additional patient with a lesion >or=6 mm and two with a lesion >or=10 mm, resulting in a sensitivity of 82% (41/50, 95% CI: 71-93%) (p = 0.50) and 72% (18/25, 95% CI: 54-90%) (p = 1.0), respectively. Specificity without CAD for polyps >or=6 mm and >or=10 mm was 84% (101/120, 95% CI: 78-91%) and 94% (136/145, 95% CI: 90-98%), respectively. With CAD, the specificity remained (nearly) unchanged: 83% (99/120, 95% CI: 76-89%) and 94% (136/145, 95% CI: 90-98%), respectively. Thus, although CTC with CAD detected a few more patients than CTC without CAD, it had no statistically significant positive influence on CTC performance.

Classifying CT image data into material fractions by a scale and rotation invariant edge model

A fully automated method is presented to classify 3-D CT data into material fractions. An analytical scale-invariant description relating the data value to derivatives around Gaussian blurred step edges--arch model--is applied to uniquely combine robustness to noise, global signal fluctuations, anisotropic scale, noncubic voxels, and ease of use via a straightforward segmentation of 3-D CT images through material fractions. Projection of noisy data value and derivatives onto the arch yields a robust alternative to the standard computed Gaussian derivatives. This results in a superior precision of the method. The arch-model parameters are derived from a small, but over-determined, set of measurements (data values and derivatives) along a path following the gradient uphill and downhill starting at an edge voxel. The model is first used to identify the expected values of the two pure materials (named L and H) and thereby classify the boundary. Second, the model is used to approximate the underlying noise-free material fractions for each noisy measurement. An iso-surface of constant material fraction accurately delineates the material boundary in the presence of noise and global signal fluctuations. This approach enables straightforward segmentation of 3-D CT images into objects of interest for computer-aided diagnosis and offers an easy tool for the design of otherwise complicated transfer functions in high-quality visualizations. The method is applied to segment a tooth volume for visualization and digital cleansing for virtual colonoscopy.

Detection of protrusions in curved folded surfaces applied to automated polyp detection in CT colonography

Over the past years many computer aided diagnosis (CAD) schemes have been presented for the detection of colonic polyps in CT Colonography. The vast majority of these methods (implicitly) model polyps as approximately spherical protrusions. Polyp shape and size varies greatly, however and is often far from spherical. We propose a shape and size invariant method to detect suspicious regions. The method works by locally deforming the colon surface until the second principal curvature is smaller than or equal to zero. The amount of deformation is a quantitative measure of the 'protrudeness'. The deformation field allows for the computation of various additional features to be used in supervised pattern recognition. It is shown how only a few features are needed to achieve 95% sensitivity at 10 false positives (FP) per dataset for polyps larger than 6 mm.

Feasibility of automated matching of supine and prone CT-colonography examinations

Matching of prone and supine positions in CT colonography may improve accuracy of polyp detection. The purpose of this study was to investigate the feasibility of automatic prone-supine matching in CT-colonography using proven polyps as fixed points of reference. The method is based on similarities in the direction of centre-lines and allows for compression and extraction of the centre-lines in both positions. To illustrate the impact of the match error of the new method in practice, the visibility of the matched polyps in a primary three-dimensional unfolded cube setting was determined as well. The method was compared with a method that relies on the normalized distance along the centre-line (NDAC method). The median absolute match error was 14 mm (range 0-59 mm, average 20 mm) either proximal or distal from the actual polyp in prone position. In the observer study, 70% (26/37) of the polyps were directly visible in prone view. The overall difference in median absolute match error between both methods was small (2 mm), although half way along the centre-line there were polyps with substantial differences in match error (larger with NDAC). We concluded that automated prone-supine matching of CT-colonography studies is feasible and has a low match error. The difference with the NDAC method was small and not significant, although half way along the centre-line some differences were seen.

Lines of curvature for polyp detection in virtual colonoscopy

Computer-aided diagnosis (CAD) is a helpful addition to laborious visual inspection for preselection of suspected colonic polyps in virtual colonoscopy. Most of the previous work on automatic polyp detection makes use of indicators based on the scalar curvature of the colon wall and can result in many false-positive detections. Our work tries to reduce the number of false-positive detections in the preselection of polyp candidates. Polyp surface shape can be characterized and visualized using lines of curvature. In this paper, we describe techniques for generating and rendering lines of curvature on surfaces and we show that these lines can be used as part of a polyp detection approach. We have adapted existing approaches on explicit triangular surface meshes, and developed a new algorithm on implicit surfaces embedded in 3D volume data. The visualization of shaded colonic surfaces can be enhanced by rendering the derived lines of curvature on these surfaces. Features strongly correlated with true-positive detections were calculated on lines of curvature and used for the polyp candidate selection. We studied the performance of these features on 5 data sets that included 331 pre-detected candidates, of which 50 sites were true polyps. The winding angle had a significant discriminating power for true-positive detections, which was demonstrated by a Wilcoxon rank sum test with p < 0.001. The median winding angle and inter-quartile range (IQR) for true polyps were 7.817 and 6.770 - 9.288 compared to 2.954 and 1.995 - 3.749 for false-positive detections.

Segmentation and size measurement of polyps in CT colonography

Virtual colonoscopy is a relatively new method for the detection of colonic polyps. Their size, which is measured from reformatted CT images, mainly determines diagnosis. We present an automatic method for measuring the polyp size. The method is based on a robust segmentation method that grows a surface patch over the entire polyp surface starting from a seed. Projection of the patch points along the polyp axis yields a 2D point set to which we fit an ellipse. The long axis of the ellipse denotes the size of the polyp. We evaluate our method by comparing the automated size measurement with those of two radiologists using scans of a colon phantom. We give data for inter-observer and intra-observer variability of radiologists and our method as well as the accuracy and precision.

Evidence and diagnostic reporting in the IHE context

Capturing clinical observations and findings during the diagnostic imaging process is increasingly becoming a critical step in diagnostic reporting. Standards developers-notably HL7 and DICOM-are making significant progress toward standards that enable exchanging clinical observations and findings among the various information systems of the healthcare enterprise. DICOM-like the HL7 Clinical Document Architecture (CDA) -uses templates and constrained, coded vocabulary (SNOMED, LOINC, etc.). Such a representation facilitates automated software recognition of findings and observations, intrapatient comparison, correlation to norms, and outcomes research. The scope of DICOM Structured Reporting (SR) includes many findings that products routinely create in digital form (measurements, computed estimates, etc.). In the Integrating the Healthcare Enterprise (IHE) framework, two Integration Profiles are defined for clinical data capture and diagnostic reporting: Evidence Document, and Simple Image and Numeric Report. This report describes these two DICOM SR-based integration profiles in the diagnostic reporting process.

Aspects of computer-aided detection (CAD) and volumetry of pulmonary nodules using multislice CT

With the superb spatial resolution of modern multislice CT scanners and their ability to complete a thoracic scan within one breath-hold, software algorithms for computer-aided detection (CAD) of pulmonary nodules are now reaching high sensitivity levels at moderate false positive rates. A number of pilot studies have shown that CAD modules can successfully find overlooked pulmonary nodules and serve as a powerful tool for diagnostic quality assurance. Equally important are tools for fast and accurate three-dimensional volume measurement of detected nodules. These allow monitoring of nodule growth between follow-up examinations for differential diagnosis and response to oncological therapy. Owing to decreasing partial volume effect, nodule volumetry is more accurate with high resolution CT data. Several studies have shown the feasibility and robustness of automated matching of corresponding nodule pairs between follow-up examinations. Fast and automated growth rate monitoring with only few reader interactions also adds to diagnostic quality assurance.

Three-dimensional display modes for CT colonography: conventional 3D virtual colonoscopy versus unfolded cube projection

The authors compared a conventional two-directional three-dimensional (3D) display for computed tomography (CT) colonography with an alternative method they developed on the basis of time efficiency and surface visibility. With the conventional technique, 3D ante- and retrograde cine loops were obtained (hereafter, conventional 3D). With the alternative method, six projections were obtained at 90 degrees viewing angles (unfolded cube display). Mean evaluation time per patient with the conventional 3D display was significantly longer than that with the unfolded cube display. With the conventional 3D method, 93.8% of the colon surface came into view; with the unfolded cube method, 99.5% of the colon surface came into view. Sensitivity and specificity were not significantly different between the two methods. Agreements between observers were kappa = 0.605 for conventional 3D display and kappa = 0.692 for unfolded cube display. Consequently, the latter method enhances the 3D endoluminal display with improved time efficiency and higher surface visibility.

Automated 3-D PDM construction from segmented images using deformable models

In recent years, several methods have been proposed for constructing statistical shape models to aid image analysis tasks by providing a priori knowledge. Examples include principal component analysis of manually or semiautomatically placed corresponding landmarks on the learning shapes [point distribution models (PDMs)], which is time consuming and subjective. However, automatically establishing surface correspondences continues to be a difficult problem. This paper presents a novel method for the automated construction of three-dimensional PDM from segmented images. Corresponding surface landmarks are established by adapting a triangulated learning shape to segmented volumetric images of the remaining shapes. The adaptation is based on a novel deformable model technique. We illustrate our approach using computed tomography data of the vertebra and the femur. We demonstrate that our method accurately represents and predicts shapes.