Polyp False positive

The polyp candidates thus detected include a large number of false-positive findings. Studies have shown that prominent folds and stool are major sources of false-positives in CAD (Yoshida et al. 2002a Yoshida et al. 2002b). Various methods characterising false-positives have been developed... 

A group at Stanford University reported a 100% sensitivity with 7.0 false-positives per data set (only the supine data set of each patient was used) based on 8 patients that included a total of 7 polyps > 10 mm in 4 patients (Paik et al. 2004). The sensitivity was less than 50% at the same false-positive rate for 11 polyps 5-9 mm that were found in 3 of the above 8 patients. A group at the National Institutes of Health reported a 90% sensitivity with 15.7 false-positives per data set, based on 40 patients (80 data sets) that included a total of 39 polyps >3 mm in 20 patients (Jerebko et al. 2003b). In a separate study, they reported 98% sensitivity with 2.1 false-positives per case (1.0 per data set) for polyps > 10 mm, and 61% with 8 false-positives per case (4 false-positives per data set) for polyps >6 mm, based on 792 patients (1,584 data sets), which were a subset of the cases from the multicenter clinical trial (Pickhardt et al. 2003) and included a total of 119 >6 mm polyps (of which 28 were 10 mm or larger) (Summers et al. 2005b). 

A group at the University College Hospital reported the performance of a commercial CAR system (ColonCAR version 1.2, MedicSight PLC, London, UK), which showed a sensitivity of 81% with 13 false-positives per patient, based on 25 patients (50 data sets), including a total of 21 polyps >5 mm in 14 patients (Taylor et al. 2006a). 

Nevertheless, the above studies appear to indicate that CAD is likely to succeed in detecting polyps with high sensitivity and a low false-positive rate - it appears that the performance of CAD schemes ranges between 70 and 100% by-patient sensitivity for polyps >6 mm, with 2-8 false-positives per patient, and the performance can reach up to 100% by-patient sensitivity with 1.3 false-positives per patient for polyps >5 mm (Nappi and Yoshida 2003). Reported performance of CTC showed that, for human readers, the pooled by-patient sensitivity for polyps > 10 mm and for those 6-9 mm was 85% and 70%, respectively (Mulhall et al. 2005). Comparing this performance with that of CAD, it appears that the performance, especially the sensitivity, of CAD is comparable or even superior to that of human readers. 

Studies have shown that most of the false-positives detected by CAD tend to exhibit polyp-like shapes, and the major causes of CAD false-positives are thickened haustral folds and retained stool (PiCKHARDT 2004 Yoshida et al. 2002a Yoshida et al. 2002b), which is similar to the cause of false-positives by human readers (Pickhardt 2004). An example of the breakdown of the false-positive sources is the following Approximately half (45%) of the false-positives are caused by folds or flexural pseudotumors. They consist of sharp folds at the sigmoid colon, folds prominent on the colonic wall, two converging folds, ends of folds in the tortuous colon, and folds in the not-well-distended colon. One fifth (20%) are caused by solid stool, which is often a major source of error for radiologists as well. Approximately 15% are caused by residual materials inside the small bowel and stomach, and 10% are... 

Fortunately, the majority of the CAD false-positives - approximately 80-90% of them - can be dismissed relatively easily based on their characteristic locations and appearance, and thus they are not a productivity hindrance (Okamura et al. 2004 Taylor et al. 2003). For example, a falsepositive detection on a thickened fold can be easily dismissed in a 3D endoluminal view, in which the reader can see the global structure of a fold on which a small bump that CAD points to is located. False-positives due to ileocecal valves and the rectal tube can easily be dismissed based on their anatomic location and shape a semi-automated recognition of ileocecal valves (Summers et al. 2004) and rectal tubes (lORDANESCU and Summers 2004 Suzuki et al. 2006) may make this already easy task even easier (Summers et al. 2004). Solid stool can be difficult false-positives to dismiss however, one may distinguish them from polyps by visual correspondence analysis between prone and supine views this relatively elaborate task can also be facilitated by a computerized correspondence between supine and prone data sets (Nappi et al. 2004c). 

Fig. 27.7.a-d Example of CAD false-positives. a Prominent fold. The tip of the fold in the figure (arrow) appears to he a cap-like structure, and thus it was incorrectly identified hy CAD as a polyp, b Solid stool. This polyp-mimicking stool has a cap-like appearance and a solid internal texture pattern, and thus it was detected incorrectly as a polyp, c Ileocecal valve. The tip of the ileocecal valve often has the cap-like appearance of a polyp and thus can be a cause of false-positives in CAD. d Residual materials inside the small bowel and stomach. Although a majority of the small bowel and stomach is removed in the colon extraction step, a small piece of them may be extracted along with the colon, and thus residual materials in the small bowel and stomach can cause false-positives. (Reprint, with permission, from (Yoshida and Dachman 2005)... 

However, there are types of false-positives, such as solid stool that mimics the shape of polyps and adheres to the colonic wall, vdiich are difficult to differentiate from polyps even for an experienced radiologist. One study has shown that approximately one false-positive per case required more detailed problem solving (Taylor et al. 2003) another study showed that, on average, 23% of the CAD false-positives were erroneously interpreted as polyps hy a reader (Okamura et al. 2004). Moreover, the pattern of the false-positives may differ across different CAD systems. More research is required for establishing how radiologists can remove these false-positives to make a correct final diagnosis more reliably. 

CAD techniques for CTC have advanced substantially during the last several years. As a result, a fundamental CAD scheme for the detection of polyps has been established, and commercial products are now available. Thus far, CAD shows the potential for detecting polyps and cancers with high sensitivity and with a clinically acceptable low false-positive rate. However, CAD for CTC needs to be improved further for more accurate and reliable detection of polyps and cancers. There are a number of technical challenges that CAD must overcome, and the resulting CAD systems should be evaluated based on large-scale, multicenter, prospective clinical trials. If the assistance in interpretation offered by CAD is shown to improve the diagnostic performance sub-... 

Nappi J, Yoshida H (2002) Automated detection of polyps with CT colonography evaluation of volumetric features for reduction of false-positive findings. Acad Radiol 9 386-397... 

Suzuki K, Yoshida H, Nappi J et al (2006) Massive-training artificial neural network (MTANN) for reduction of false-positives in computer-aided detection of polyps suppression of rectal tubes. Medical Physics 10 3814-3824 Taylor SA, Halligan S, Burling D et al (2006a) Computer-assisted reader software versus expert reviewers for polyp detection on CT colonography. AJR Am J Roentgenol 186 696-702... 

The primary aim of CT colonography is accurate identification of significant colorectal polyps and cancers in a minimally invasive manner. For CT colonography to be a safe, accurate and attractive alternative to colonoscopy, radiologists reading these studies must confidently recognise polyps and cancers, identify pitfalls and therefore reduce the number of false positive findings, and report... 

One of the most common colorectal morphologies is the focal polypoid lesion. This is also the most common morphology of the false positive lesion of retained stool. Thus discernment between a focal polyp and stool are critical. Key features include the following ... 

Fig.7.3a,b. Polypoid lesions with focal pockets of air seen in true polyp vs stool, best shown in axial 2D MPR a true positive sessile polyp (arrow) with air around edges of lesion (arrowheads), where lesion abuts the wall b false positive of stool (arrow) with central pockets of air (arrowheads)... 

Fig. 7.4a-d. False positive lesion of retained pill (arrow) with characteristic shift of position a axial 2D MPR supine image (W 1500, L -200) demonstrates a polypoid lesion b axial 2D MPR view in soft tissue window settings (W 400, L 10) better shows low density of pill c axial 2D MRP prone image in soft tissue settings better demonstrates shift to dependent position of pill, consistent with false positive d 3D volume rendered view of pill mimics a polyp... 

Fig. 8.29a,b. False positive diagnosis complex folds a axial image shows the splenic flexure, with a thickened nodular-like fold (arrow) b corresponding endoluminal 3D image clearly shows that the thickened nodular appearance is to be explained by the complexity of the folds at the splenic flexure. Lesson Complex or thickened folds are typically encountered at the splenic and hepatic flexures, and should be differentiated from sessile cancers or polyps. Endoluminal 3D images are extremely helpful for differential diagnosis. Compare with Fig. 8.18... 

VAN Gelder et al. (2004c) compared primary 2D evaluation with a primary 3D evaluation method (unfolded cube projection) in a series of 77 patients. Mean sensitivity for large polyps for the primary 3D and 2D review methods were 83 and 72%, respectively. The specificity was 92 and 94% respectively. Fewer perceptive errors, although not statistically significant (p=0.06), were made with the primary 3D method than with the primary 2D method although at expense of a slight increase of the number of false positives. 

To date, most of the CAD schemes developed in academia and in industry comprise of the following four fundamental steps (1) extraction of the colonic wall from the CTC images, (2) detection of polyp candidates in the extracted colon, (3) characterization of false positives, and (4) discrimination between false positives and polyps. A brief descrip-... 

The final detected polyps are obtained by application of a statistical classifier based on the image features to the differentiation of polyps from false positives. Investigators use parametric classifiers such as quadratic discriminant analysis (Yoshida and Nappi 2001), non-parametric classifiers such as artificial neural networks (Jerebko et al. 2003b Kiss et al. 2002 Nappi et al. 2004b), a committee of neural networks (Jerebko et al. 2003a), and a support vector machine (Gokturk et al. 2001). In principle, any combination of features and a classifier that provides a high classification performance should be sufficient for the differentiation task. 

Among the studies published in peer-reviewed journals that describe a full CAD scheme, the CAD scheme developed at the University of Chicago yielded a 95% by-polyp sensitivity, with an average of 1.5 false positives per patient (0.7 false positives per... 

The CAD system at the University Hospital Gast-huisberg achieved an 80% hy-polyp sensitivity, with 8.2 false positives per patient (4.1 false positives per data set), based on 18 patients, with 15 polyps 5 mm in 9 patients (Kiss et al. 2002). In this study, fecal tagging was used for most of the cases. A group at Stanford reported a 100% sensitivity with 7.0 false positives per data set (only the supine data set of each patient was used) based on 8 patients that included a total of 7 polyps >10 mm in 4 patients (Pair et al. 

---