What can be expected from computerized image-processing techniques in the quantification and mapping of Barrett's mucosa?
R. Kim, J.C. Reynolds (Pittsburgh)
Barrett's esophagus (BE) is the most common premalignant lesion of the upper gastrointestinal tract. It is associated with a 30-125 fold increased risk of developing esophageal adenocarcinoma [1-3]. The alarming and epidemic increase in the incidence of esophageal adenocarcinoma in the United States has generated significant interest into the study of it's pathogenesis [4-6]. Therefore, significant interest has been directed towards the characterization and treatment of BE. While there are currently no therapies available that reliably result in complete regression of BE, several promising treatments cause partial regression [7-11].
The risk of malignant transformation in BE should be related in part to the extent of it's mucosal involvement [12-16]. The current standards for estimating the extent of BE involve measuring the length of columnar lined tissue in the tubular esophagus. Length has been defined as the distance of the height of the most proximal columnar epithelium from the gastroesophageal junction (GEJ) [17-20]. While in some patients this proximally relocated squamo-columnar junction (SCJ) occurs as a circumferential band, the majority of BE patients have an irregular SCJ with tongues of metaplastic tissue of varying lengths emanating from the GEJ [21-22]. This undulating border of squamous and columnar epithelium is made even more irregular by the presence of islands of both squamous and metaplastic tissue interspersed throughout the esophagus. Recent data suggest that the therapeutic response to high-dose proton pump inhibitor therapy in BE patients promotes the development of squamous islands irregularly distributed in the metaplastic epithelium, with little to no change in overall length [10, 11]. Therefore, a considerable change in the extent of mucosa covered by columnar epithelium can occur without an observed change in the "length of involvement". On the other hand, the highest peak of a tongue can recede and be reported as improvement, while the overall area of BE increases as a consequence of covering the squamous islands and filling in the valleys of normal mucosa. Given this irregular shape of BE and its tendency to heal by developing squamous islands, the use of length for evaluating the extent and the interval change in the extent of disease over time or with therapy has major conceptual limitations. Regrettably, every previous attempt to quantitatively study the natural history of BE or its response to therapy has used length as a measure of change.
A novel computer imaging analysis software program has been developed in our laboratory that corrects for the distortion of the wide angle endoscopic lens and creates accurate two-dimensional color maps from endoscopic photographs . The following text will demonstrate the applicability of this technology in calculating the area of BE.
Software was developed on an Apple Macintosh computer (Apple Computer, Inc., Cupertino, CA 95014-6299) and has been previously described . The computer first corrects for the distortion of the endoscopic lens and calculates the center of the circular endoscopic photograph. The program then recreates a three-dimensional cylindrical image and "unrolls" it into a planar image (Figure 1).
Figure 1. Schematic of computer transformation of endoscopic photographs into two-dimensional images. 1. Perspective view of the cylindrical esophagus. 2. Endoscopic view as seen and photographed through the endoscope. The computer program will utilize this image, correct for the distortion of the lens, and transform it back into the three-dimensional cylinder as seen in 1. 3. The computer then mathematically cuts and "unrolls" the cylinder transforming it into a two-dimensional image as seen here.
The technique for photographing and calculating the area of BE is described below. Endoscopy is performed with either Pentax, Fujinon or Olympus video endoscopes. Photography of the esophagus is obtained in a systematic way such that the lumen is centered on the video screen, and all photographs are taken at end-expiration with maximal insufflation.
These digitized endoscopic images are loaded onto a Macintosh computer (Apple Computer, Inc., Cupertino, CA 95014-6299) in a Targa format. Once loaded, the images are opened into a Photoshop program (Adobe Systems Inc., Mountain View, CA 94039-7900) and converted to an 8-bit PICT file. These PICT files are then automatically transformed into "unrolled" planar images and stacked by the Quantitative Endoscopic Imaging program, producing a two-dimensional map of the photographed esophagus. The perimeter of the metaplastic tissue is manually traced using the computer tools and the area under the curve is calculated. Squamous islands are then manually traced and this area automatically subtracted from the overall area.
The preliminary in vitro testing of this software was performed on models of BE . Models consisted of drawings of BE rolled up and placed in plastic tubes. These models were then endoscoped and photographed as the endoscope was withdrawn from the base of the tube. Computer generated two-dimensional maps of these models were then created utilizing our software (Figure 2). The area of "metaplastic" tissue from the computer generated maps was compared to the actual area of the model's "metaplastic" surface. The calculated area from the maps correlated with actual areas (r = 0.96) with an overall error of 5.2%. Color, size, shape, or diameter of the model, or the experience of the endoscopist did not affect the accuracy. Accuracy did improve by decreasing the interval between photographs from 4 cm (10.0% error) to 2 cm (4.8% error).
In human subjects, a new set of variables affecting the reliability of this endoscopic imaging technique was introduced. These included:
1) variations of measurements based on the degree of flexion or extension of the patient's neck during photography;
2) variations in the position of the esophagus due to respiration;
3) variations of luminal size secondary to esophageal motility;
4) variations in the determination of the GEJ;
5) tortuosity of the human esophagus.
In order to minimize these potential sources of error, standard methods of photographing the esophagus were used including maximum neck flexion, the use of a standard bite block, and obtaining photographs only during maximal distention at end expiration.
Twenty BE patients were initially studied . In order to determine if the technique was reproducible between endoscopists, each patient was photographed by two independent endoscopists during the same session. The calculated area of BE from the maps produced from the photographs correlated precisely when compared between the
2 endoscopists (r = 0.99). In order to determine if the technique of calculating the area of BE from endoscopic photographs was reliable, independent technicians utilized the endoscopic images to create the maps and calculate the area of BE. The inter-observer variability in area calculation was minimal between independent technicians (r = 0.99). Maps were made within 5 minutes and of sufficient clarity to allow for mapping of biopsy sites (Figure 3).
Figure 2. Example of a computer generated map created from endoscopic photographs of a Barrett's model. A. The cartoon that was used to make the model by rolling it up into a cylinder. B. Endoscopic photographs of the model taken at 2 cm intervals as the endoscope was withdrawn from the base of the cylinder. C. The two-dimensional map was created by stacking the computer generated images (transformed from the endoscopic photographs) in sequence. (See page VII for colour figure.)
In order to determine if the precision of this measure of area would be sustained if the photographic data were collected on different days, seventeen patients had area calculations performed on serial endoscopy, separated by a short enough interval that no major changes in the extent of disease would be expected . In addition, an assessment of the progression or regression of BE was made by both length and area measures and the concordance rates of these two measures determined. Area calculations from the two endoscopies were highly correlated (r = 0.99). While length measures were also well correlated (r = 0.90), the mean change in absolute length, 1.4 ± 0.2 cm, was greater than the change in absolute area (4.5 ± 1.4 cm2, equivalent to a length of 0.67 ± 0.2 cm, p = 0.001). The percent change in absolute length (26.9%) was greater than the percent change in absolute area (16%, p = 0.001). When the change in area from one endoscopy to the next was compared, no significant variation from no change was found (p = 0.39). In contrast, the differences in length measures from one endoscopy to the next did show a significant change in measures of the extent of BE (p = 0.008). This signified that there was less variation in area calculations from one endoscopy to the next when compared to length measures. Perhaps more importantly, there was discordance in the estimates of progression/regression between area and length in nine patients. The image technique detected squamous islands in 6 patients with areas as small as 0.3 cm2, and there was no change in the area of these squamous islands between endoscopies.
Figure 3. Example of a computer generated map created from endoscopic photographs of a patient with Barrett's esophagus. A. Endoscopic photographs from the patient taken at 1 cm intervals starting at the GEJ and proceeding proximally. B. The 2-dimensional map is created by stacking the computer generated images transformed in sequence. (See page VIII for colour figure.)
The incidence of adenocarcinoma of the esophagus and GEJ is increasing in epidemic proportions, and yet nearly all develop in an easily recognized, potentially treatable precursor, BE [24-32]. There is an urgent need, therefore, to develop effective treatments of both the extent and malignant potential of BE. To determine if proposed treatments are effective, accurate means of assessing the response must be developed. The inadequacies of utilizing length as measurement of the extent of BE have been discussed above.
Our studies have shown that the use of image analysis computer technology provides an accurate and reproducible calculation of the surface area of BE . Using this technology, area measures obtained from serial endoscopies are extremely well correlated . Furthermore, the absolute change in area calculation on serial endoscopy is more precise than the absolute change in length. Perhaps more important than the absolute change in extent of BE is the relative change: there is much less variability in the calculation of area than in measure of length on serial endoscopies, which significantly influences the assessment of the interval change in the extent of BE. Measures of the extent of disease as length also ignores the change in the true extent represented by squamous islands. If these islands are confirmed to be important in the healing process of BE, area measurements that incorporate the surface area covered by squamous tissue would be vital.
The role of the extent of disease of BE remains controversial. While several studies have suggested that longer segments of BE pose a higher risk of malignant transformation [12-16], evidence that adenocarcinoma may arise from short segments of BE is accumulating [26, 27]. Carcinogenesis from premalignant tissue seems to proceed through a series of genetic alterations affecting both oncogenes and tumor suppressor genes. It is biologically plausible that the risk of developing these genetic alterations should increase as the number of cells predisposed to these changes increases. Certainly, other potential factors play a major role in the malignant transformation of BE. Due to an incomplete understanding of these putative risk factors, efforts to eliminate the premalignant potential of BE have focused primarily on reducing it's extent, thereby, reducing the number of cells at risk. Clearly, the malignant potential of BE will not be eliminated unless all of the premalignant tissue is removed. However, as complete regression is likely to take years of therapy, the identification of intermediate endpoints of therapy become increasingly important.
The ability of our novel image analysis system to detect small squamous islands provides further evidence that this technique has sensitivity enough to detect even subtle changes in mucosal appearance. Therefore, this technology can be utilized to map endoscopic biopsy sites. Despite being a common premalignant condition, the progression of subtypes of BE, the role of oncogenes, regulatory genes, or genetic make-up is relatively unknown. These factors are undoubtedly more important determinants of the premalignant potential of the lesion than the extent of BE. Therefore, any study of the natural history or response to therapy must also attempt to evaluate the longitudinal progression of these factors at specific regions of the epithelium. The underlying histologic pattern of BE exists in a very complex "mosaic" pattern with dysplastic tissue interspersed among nondysplastic tissue [28, 29]. Endoscopically, however, the mucosa appears surprisingly homogenous. Small foci of carcinoma in situ or even invasive carcinoma may also exist without endoscopic identification . Therefore, surveillance biopsies of BE are obtained in a random fashion [21, 31]. These biopsies are generally lumped together in several formalin-filled jars as there is no current method available to precisely determine the site of each biopsy. This method is, however, associated with significant sampling error, which becomes immediately apparent when one biopsy reveals an indeterminate grade of dysplasia. It would be particularly important to re-evaluate these indeterminate changes by re-biopsying the same area. Therefore, the ability to accurately re-biopsy a particular site in the featureless esophagus will be crucial for the scientific study of BE.
The major limitation to this technique remains the identification of the GEJ. In our studies, the proximal border of the gastric rugae was used as an endoscopic landmark. This landmark has been shown to be an accurate estimate of the muscular junction of the tubular esophagus and stomach sac . It has been suggested that it is a fixed, reproducible anatomic landmark in patients with BE . Our previous work has shown that 2 endoscopists independently agreed upon the site of this landmark in all cases when measurements were taken at the same endoscopy . It should be recognized, however, that the use of this landmark may overestimate the extent of BE as the normal SCJ may be located up to 2 cm from the gastric rugae. Therefore, histological identification of specialized columnar epithelium is necessary to confirm the presence of BE in this area.
We conclude that computer image analysis technology provides a precise and reproducible technique for detecting changes in the extent of BE that occur with time or in response to therapy. This technological advance provides the only known method for performing area calculations, and should be considered a necessary prerequisite for the accurate determination of the extent of BE. Future enhancements in the software include the development of a platform and format independent program that can be incorporated directly into the endoscopic computer. This would significantly reduce the number of steps involved in producing the computer generated maps and, therefore, reduce the map production time. Future applicability in biopsy site mapping is currently being investigated and holds promise for providing a means for scientists to utilize BE as a novel in vivo, human model of carcinogenesis.
10. Gore S, Healey CJ, Sutton R, Eyre-Brook IA, Gear MWL, Shepherd NA, Wilkinson SP. Regression of columnar lined (Barrett's) oesophagus with continuous omeprazole therapy. Aliment Pharmacol Ther 1993;7:623-628.
13. Baneriee B, Salian N, Waits W, et al. Risk factors for developing dysplasia of adenocarcinoma in Barrett's esophagus (BE) with implications for endoscopic surveillance. Gastroenterology 1994;106:A45.
16. Menke-Pluymers MBE, Hop WC, Dees J, van Blankenstein, M, Tilanus HW, The Rotterdam Esophageal Tumor Study Group. Risk factors for the development of an adenocarcinoma in columnar-lined (Barrett's) esophagus. Cancer 1993;72(4):1155-1158.
18. Herlihy KJ, Orlando RC, Bryson JC, Bozymski EM, Carney CN, Powell DW. Barrett's esophagus: clinical, endoscopic, histologic, manometric, and electrical potential difference characteristics. Gastroenterology 1984;86:436-443.
23. Kim R, Baggott BB, Rose S, Shar AO, Mallory DL, Lasky SS, Kressloff M, Faccenda LY, Reynolds JC. Quantitative endoscopy: precise computerized measurement of metaplastic epithelial surface area in Barrett's esophagus. Gastroenterology 1995;108(2):360-366.
25. Kim R, Rose S, Shar AO, Weiner M, Reynolds JC. Extent of Barrett's metaplasia: a prospective study of the serial change in area of Barrett's measured by quantitative endoscopic imaging. Gastroint Endosc 1996;43:429.
30. Reid BJ, Weinstein WM, Lewin KJ, Haggitt RC, VanDeventer G, DenBesten L, Rubin CE. Endoscopic biopsy can detect high-grade dysplasia or early adenocarcinoma in Barrett's esophagus without grossly recognizable neoplastic lesions. Gastro 1988:94:81-90.
31. Levine DS, Haggitt RC, Blount PL, Rabinovitch PS, Rusch VW, Reid BJ. An endoscopic biopsy protocol can differentiate high-grade dysplasia from early adenocarcinoma in Barrett's' esophagus. Gastro 1993;105:40-50.