Linked color imaging application for improving the endoscopic diagnosis accuracy: a pilot study

Endoscopy has been widely used in diagnosing gastrointestinal mucosal lesions. However, there are still lack of objective endoscopic criteria. Linked color imaging (LCI) is newly developed endoscopic technique which enhances color contrast. Thus, we investigated the clinical application of LCI and further analyzed pixel brightness for RGB color model. All the lesions were observed by white light endoscopy (WLE), LCI and blue laser imaging (BLI). Matlab software was used to calculate pixel brightness for red (R), green (G) and blue color (B). Of the endoscopic images for lesions, LCI had significantly higher R compared with BLI but higher G compared with WLE (all P < 0.05). R/(G + B) was significantly different among 3 techniques and qualified as a composite LCI marker. Our correlation analysis of endoscopic diagnosis with pathology revealed that LCI was quite consistent with pathological diagnosis (P = 0.000) and the color could predict certain kinds of lesions. ROC curve demonstrated at the cutoff of R/(G+B) = 0.646, the area under curve was 0.646, and the sensitivity and specificity was 0.514 and 0.773. Taken together, LCI could improve efficiency and accuracy of diagnosing gastrointestinal mucosal lesions and benefit target biopsy. R/(G + B) based on pixel brightness may be introduced as a objective criterion for evaluating endoscopic images.


Pixel brightness of LCI images was different from that of WLE and BLI images.
In order to deepen the current understanding on the properties of endoscopic imaging, we analyzed the specific color of the images by calculating the pixel brightness for red, blue and green color, respectively (Figs 1 and 2). A total of 44 paired WLE, LCI and BLI images for the same lesion or normal mucosa were selected. Our results demonstrated that R, G and B of LCI images were higher than those of WLE and BLI images (all P < 0.05) for normal mucosa, which could validate the image enhancement by BLI and LCI technique. R of LCI images for the lesion was higher than that of BLI images (202.973 ± 26.348 vs. 198.329 ± 25.376, P = 0.000), and G of LCI images was higher than that of WLE images (128.360 ± 24.553 vs. 104.605 ± 17.974, P = 0.000), which was consistent with the design of BLI and LCI 9 . Although B of LCI images was significantly different from that of WLE and BLI images, B for normal mucosa was insignificantly from that for the lesion in all 3 endoscopic modes. Thus, we finally introduced the value of R/(G+ B) as a composite marker for evaluating the endoscopic images, which was remarkably different (LCI vs. WLE, P = 0.000; LCI vs. BLI, P = 0.002) ( Table 2). LCI imaging could facilitate the target biopsy under endoscopy. The clinical application of LCI imaging were further expanded. We correlated the endoscopic diagnosis with the pathological analysis and it was proved that LCI was most consistent with the pathology (BLI vs. WLE P = 0.029; LCI vs. WLE P = 0.000; Table 3). These results indicated that endoscopic LCI imaging could improve the efficiency and accuracy of target biopsy, which will definitely help avoid the misdiagnosis. The main color characteristics of certain lesions by LCI were also summarized (Fig. 1). Lesions of H. pylori (HP) infection were in diffusion red color (n = 20), of inflammation or cancer were in red color (n = 33), of intestinal metaplasia were in purple color (n = 21), of atrophy were in white color (n = 11) and normal mucosa were manifested in yellow color (n = 20). These data support that specific color feature of LCI images might be closely correlated with the pathology.

Certain pixel brightness of LCI images could differentiate the lesions from normal mucosa.
Although the endoscopy is useful in diagnosing mucosal lesion of gastrointestinal tract, most of the lesions were subjectively identified and there was still no objective judgment on the specific endoscopic images. Thus, based on our data, we selected R/(G + B) as a potential marker for endoscopically differentiating the lesions from normal mucosa. A total of 116 LCI images were extracted from our computerized database. ROC curve for R, G, B and R/(G + B) was drawn, respectively (Fig. 3). The area under curve for R/(G+ B) was 0.646 (P = 0.008). At the cutoff of 0.8020, the sensitivity and specificity was 0.514 and 0.773 (Table 4). Taken together, pixel brightness might serve as a new method to analyze the features of the endoscopic images for proper diagnosis, especially for chromoendoscopy which is focused on the color change.

Discussion
Great progress has been made on the endoscopic technique in order to improve the diagnostic accuracy and therapeutic efficacy 10,11 . Among them, chromoendoscopy and magnified endoscopy have the greatest contribution, which could help identify the lesions and observe the mucosa's microstructure 12  based on BLI technique, which is a chromoendoscopic technique via a specially designed laser system 7 . LCI is designed to be able to obtain a sufficient brightness and easily differentiated contrast in hue. Furthermore, LCI and BLI put more emphasis on the color change of the mucosa. Our study evaluated the clinical application of LCI in diagnosing gastrointestinal mucosal lesions and the main characteristics of the LCI images were summarized for the first time, which will definitely optimize the current clinical strategy for such patients in future.
There is no doubt that the endoscopic diagnosis has been subjectively made by the endoscopists with quite limited objectivity 13 . In addition, no efficient and accurate endoscopic criteria or markers have been proposed by previous studies 10,14,15 . Thus, we used the value of pixel brightness in RGB color model to evaluate the color of the endoscopic images and further analyzed its correlations with certain clinical data. RGB color model is an additive model of a series of colors that could be reproduced by red, green and blue light in different arrangements. Pixel brightness is a physical unit in a image, which is widely used for analyzing digital imaging as a continuous variable, but pixel brightness for RGB model has not been ever reported in evaluating endoscopic images [16][17][18] . Thus, we hypothesized that pixel brightness for RGB color may be applied to judging the endoscopic images, which is of translational significance. Our results found that LCI images for the lesions had higher R than BLI but lower G compared with WLE, which could be explained by the imaging principle of LCI. LCI was BLI with red light added, and thus green and blue color in BLI will be purple and yellow color, respectively. Since the B for the lesion was insignificantly different from that for normal mucosa observed by WLE, LCI and BLI, B was excluded for further analysis. The wavelength for green (577-492 nm) and blue light (450-435 nm) is approaching, while the wavelength for red light is 760-622 nm. The wavelength determines the properties of certain light including the ability of tissue penetration. The longer the light wavelength, the stronger the tissue penetration 19,20 . Seen from this, the value of R/(G + B) is a potential composite marker for investigating the physical properties of WLE, LCI and BLI technique, which was also supported by our findings that R/(G + B) of LCI images for the lesions were greatly higher than those of WLE and BLI, respectively (P < 0.05). These data also indicated that the calculation of pixel brightness for RGB color may be a useful method of evaluating the endoscopic images.
Target biopsy and real-time observation has been highly recommended in recent years 21,22 . We compared the diagnostic accuracy of WLE, LCI and BLI using the pathology as the gold standard. It was observed that LCI endoscopic diagnosis was significantly correlated with the pathology (P = 0.000), and different lesions possessed    Table 3. The LCI imaging could benefit the endoscopic target biopsy.
specific color features. In our cohort, it was also supposed that the color of LCI images could be also used for evaluating the tumor infiltration. Lesions of early cancer usually were presented as red area with yellow color, while those of advanced cancer were as red area with white color. Furthermore, we focused on the identification of a quantifiable marker for obtaining endoscopic diagnosis. ROC curve for differentiating the lesion from normal mucosa was calculated and it revealed that R/(G + B) may be a potential endoscopic marker. The area under the curve was not quite large with medium sensitivity and specificity, which may be associated with the fact that the endoscopic images are manually recorded and thus exposure delay and a small shake may impair the quality of the images. Our study could establish a new computerized endoscopic diagnostic model as Miyaki et al. proposed 23 , which could enlighten the future research.
Our study also has limitations. First, this pilot study was a retrospective analysis and clinical data were retrieved from our electronic database. There were 10 patients who were excluded due to incomplete information. Second, the sample size was not quite large and only 44 patients were enrolled. However, as a pilot study, our results could be able to provide relatively reliable evidence for elaborating the valuable clinical usage of LCI in gastrointestinal endoscopic examinations. Third, our study mainly focused on the role of RGB color pixel brightness and other factors associated with color enhancement have not been fully examined. A multicenter large-scale prospective investigation on the clinical application of LCI technique will be conducted soon.
Gastrointestinal endoscopy has incomparable advantages over other diagnostic examinations like ultrasound and radiological images 24 . The development of LCI may be a new milestone in the history of endoscopy, which could guarantee relatively high diagnostic efficiency and accuracy by modulating the color of the imaging. We also demonstrated that pixel brightness for RGB color model could potentially quantifying the endoscopic images as a objective marker, which will be validated in future investigations. New quantifiable endoscopic criteria for diagnosing certain mucosal lesions should be explored.   was approved by the Ethic Committee of the 307 Hospital of Academy of Military Medical Science in accordance with Declaration of Helsinki. All the methods were performed following the relevant guidelines and regulations.

Patients.
Endoscopy and LCI. All the endoscopic examinations were completed by experienced endoscopists. The indications were listed in Table 1. EC-L590ZW endoscope with the LASEREO system (FUJIFILM Co., Tokyo, Japan) was used and all the endoscopic procedures were performed routinely. No patients were under anesthesia. There were 17 patients who underwent colonoscopy and 37 who underwent gastroduodenoscopy. During the endoscopic examination, all the lesions and surrounding normal mucosa were observed by white light (WLE), LCI and BLI in turns and typical images were recorded, respectively (Fig. 1). Biopsy was performed for each lesion and the pathological diagnosis were collected. If more than 1 lesion was detected, the most severe one was included for further analysis.
Image analysis. MatLab software (USA) was applied to analyze the WLE, LCI and BLI images as the previous study described 25 . The brightness of all the endoscopic images has been set to be comparable. The area of interest in the endoscopic images was selected, and the pixel brightness for red (R), green (G) and blue (B) color was evaluated (Fig. 2). The value of R/(G + B) was also introduced as a composite marker for comprehensive analysis.
Statistical analysis. All the statistical analysis were performed using SPSS 17.0 software. Continuous and categorical data were shown as mean ± standard deviation (SD) and percentage (%), respectively. The differences on continuous data were tested by student t-test and on categorical data by chi-square test. Receiver operating characteristic (ROC) curve of R, G, B and R/(G + B) for differentiating the lesion from normal mucosa was analyzed and the area under curve, cutoff, sensitivity and specificity were also calculated. Two-tailed P value less than 0.05 was considered as statistically significant.