https://orcid.org/0000-0001-6647-6325
Figures
Abstract
Indocyanine green (ICG) has been used in clinical practice for more than 40 years and its safety and preferential accumulation in tumors has been reported for various tumor types, including colon cancer. However, reports on clinical assessments of ICG-based molecular endoscopy imaging for precancerous lesions are scarce. We determined visualization ability of ICG fluorescence endoscopy in colitis-associated colon cancer using 30 lesions from an azoxymethane/dextran sulfate sodium (AOM/DSS) mouse model and 16 colon cancer patient tissue-samples. With a total of 60 images (optical, fluorescence) obtained during endoscopy observation of mouse colon cancer, we used deep learning network to predict four classes (Normal, Dysplasia, Adenoma, and Carcinoma) of colorectal cancer development. ICG could detect 100% of carcinoma, 90% of adenoma, and 57% of dysplasia, with little background signal at 30 min after injection via real-time fluorescence endoscopy. Correlation analysis with immunohistochemistry revealed a positive correlation of ICG with inducible nitric oxide synthase (iNOS; r > 0.5). Increased expression of iNOS resulted in increased levels of cellular nitric oxide in cancer cells compared to that in normal cells, which was related to the inhibition of drug efflux via the ABCB1 transporter down-regulation resulting in delayed retention of intracellular ICG. With artificial intelligence training, the accuracy of image classification into four classes using data sets, such as fluorescence, optical, and fluorescence/optical images was assessed. Fluorescence images obtained the highest accuracy (AUC of 0.8125) than optical and fluorescence/optical images (AUC of 0.75 and 0.6667, respectively). These findings highlight the clinical feasibility of ICG as a detector of precancerous lesions in real-time fluorescence endoscopy with artificial intelligence training and suggest that the mechanism of ICG retention in cancer cells is related to intracellular nitric oxide concentration.
Citation: Kim J, Kim H, Yoon YS, Kim CW, Hong S-M, Kim S, et al. (2023) Investigation of artificial intelligence integrated fluorescence endoscopy image analysis with indocyanine green for interpretation of precancerous lesions in colon cancer. PLoS ONE 18(5): e0286189. https://doi.org/10.1371/journal.pone.0286189
Editor: Shiki Fujino, Osaka International Cancer Institute: Osaka Kokusai Gan Center, JAPAN
Received: February 17, 2023; Accepted: May 11, 2023; Published: May 25, 2023
Copyright: © 2023 Kim et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the paper and its Supporting Information files.
Funding: This research was supported by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant numbers: HI22C1721 and HR21C0198); and Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (grant numbers: 2021R1A6A1A03040260 and 2022R1F1A1068465). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Colorectal cancer (CRC) is the most common cancer worldwide and the leading cause of cancer-related deaths [1]. The progression of adenoma leads to CRC, and involves the conventional tubular adenoma and alternative serrated polyp pathways [2]. Chronic inflammation is a major risk factor that induces different morphologies and dysplasia-carcinoma progression in patients with inflammatory bowel disease (IBD) [3]. Hence, the detection of early phase cancer or dysplasia as precancerous lesions could lead to decreased mortality.
Currently, white-light endoscopy is the gold standard for screening colonoscopy; however, the miss rate of adenomas during endoscopy is 24.1% and for smaller and flat or sessile polyps it is significantly higher [4]. For sensitive diagnosis of CRC lesions, advanced molecular imaging techniques, such as fluorescence endoscopy, which targets tumor-specific molecules, have been developed [5–7]. Fluorescence-based tumor imaging can be a useful technique for early cancer detection owing to its advantageous high specificity and sensitivity, real-time imaging, and absence of radiation exposure [8]. In previous studies, a fluorescence-labeled compound, which could bind to tumor-specific molecules, such as vascular endothelial growth factor (VEGF) [9, 10], epidermal growth factor receptor (EGFR) [11], colon cancer specific protein-2 (CCSP-2) [12], endothelial A receptor (ETAR) [13], and c-Met [14] was developed to detect colon cancer. Although various probes have been used, their safety in the human body and their ability to detect precancerous lesions, such as adenoma, serrated polyps, and dysplasia, must be confirmed before applying this technology in clinical endoscopic examination [15].
Indocyanine green (ICG) is a tricarbocyanine iodide dye molecule that is amphiphilic, relatively nontoxic fluorescent contrast agent exhibiting a good safety profile for hepatic function examinations [16, 17] and has been approved for clinical applications in the USA and in some European and Asian countries. It is known that intravenously injected ICG preferentially accumulated in malignant tissues than normal through its binding to macromolecular serum proteins or cellular endocytosis [18–22]. Based on the preferred chemical properties for in vivo imaging, various clinical and preclinical applications of ICG have been reported, including tumor detection in the liver [23, 24], breast [25], head and neck [26], lungs [27], ovaries [28], and colon [29, 30] of humans. Despite little background noise in the gastrointestinal tract under near infrared light exposure with ICG fluorescence endoscopy [31], verification of the practicality of ICG injections to detect colon cancer and precancerous lesions in the clinic setting has been insufficient.
Artificial intelligence (AI) has been widely applied in medicine, including colon cancer diagnosis and treatment. Recently, several authors have presented research on the application of AI in early diagnosis of colon cancer using convolutional neural networks (CNNs), a type of deep learning algorithm [32, 33]. In these studies, the possibility of the clinical application of AI was demonstrated by using white-light colon endoscopy, tomography (PET-CT), and histopathology images [34–36].
In the present study, we investigated the correlation between fluorescence intensity of ICG-based fluorescence endoscopy and colon cancer specific targets in a colitis-associated colon cancer model. To verify the feasibility of ICG for clinical endoscopy, we evaluated its ability to detect precancerous lesions by conducting AI analysis of ICG fluorescence endoscopy images. This successful trial of an AI integrated ICG fluorescence endoscopy image analysis demonstrated relevance in the interpretation of precancerous lesions of colon cancer patients. Furthermore, we evaluated the mechanism of ICG retention in tumor cells and its relation with inducible nitric oxide synthase (iNOS) expression using a cell culture experiment. Finally, we validated the correlation between fluorescence intensity and malignant lesions in patient tissues.
Materials and methods
Patient consent for human participants
All patient tissue samples and data used in this study were provided by the Institutional Review Board of Asan Medical Center (IRB; protocol No. 2017–0837). Comprehensive approvals for basic research were obtained from all patients, and informed consent was obtained by written from all patients and/or their legal guardian(s). This study was conducted in accordance with the Ethical Guidance of the Declaration of Helsinki.
Animal study ethics
All animal studies were approved by Institutional Animal Care and Use Committee of the Asan Institute for Life Sciences, Asan Medical Center (IACUC; Approval No. 2019-12-205). All animal care and experimental procedures aligned with appropriate instructions and regulations of the Institutional Animal Care and Use Committee (IACUC) at Asan Medical Center and ARRIVE guidelines.
Cell lines
The human colon cancer cell lines, HCT116, SW480, and CCD841 were obtained from the American Type Culture Collection (ATCC; Manassas, VA, USA). Cells were maintained in Dulbecco’s Modified Eagle Medium/High Glucose (DMEM/High Glucose; Gibco; Thermo Fisher Scientific, Inc., Waltham, MA, USA) supplemented with 10% fetal bovine serum (Hyclone, Logan, UT, USA) and 1% Antibiotic-Antimycotic (Gibco; Thermo Fisher Scientific, Inc.). The cell lines were grown at 37°C in a humidified incubator containing 5% CO2.
Establishment of the azoxymethane/dextran sulfate sodium (AOM/DSS)-induced colitis-associated colon cancer in a mouse model
15 Male Balb/c mice (age, 6–7-week-old; weight, 20–23 g) were purchased from Orient Bio (Seongnam, South Korea). All animals were kept in polycarbonate cages (3 mice per cage) under controlled environmental conditions at 22°C and 50% humidity in a 12 hours light/dark cycle. Enough food and water were made available to the animals. Starting from 7 to 8 weeks old, mice were administered a single intraperitoneal injection of azoxymethane (AOM) (10 mg/kg body weight; Wako Pure Chemical Co., Osaka, Japan). One week after the AOM injection, colitis was induced by two intermittent one-week-cycle administration of 2% dextran sodium sulfate (DSS; MP Biochemicals, Santa Ana, CA, USA; MW 36,000–50,000) in drinking water. Mice that showed critical weight loss during the first week of DSS administration had their DSS treatment discontinued immediately. One week after discontinued DSS administration, tumor progression was confirmed weekly through optical/fluorescent colonoscopy (Vetcom; Karl Storz, Tuttlingen, Germany). 10~14 weeks after AOM administration, all mice were sacrificed by inhalation anesthesia (isoflurane; Terrell, Piramal Critical Care, PA, USA), excised, and opened longitudinally to remove their colons (S1 Fig in S1 File). Colons were cleaned with PBS and number of polyps from the start of cecum to the anus. Three sections (~2 cm) of rectum samples were recovered from each mouse, fixed in 4% formalin embedded in paraffin or flash-frozen in liquid nitrogen and stored at −80°C until analyzed for histology. All animals were kept in polycarbonate cages (3 mice per cage) under controlled environmental conditions at 22°C and 50% humidity in a 12 hours light/dark cycle. Enough food and water were made available to the animals. No animals were found dead and any mice were not reached endpoint criteria written in approved Animal Care and Use Protocols (IACUC).
Orthotopic colonic submucosal implantation of CRC cells
Before the orthotopic submucosal injection was performed, a fabricated needle was prepared for the working channel of the colonoscopy (Vetcom; Karl Storz). A 30G needle was bound to a 23G needle with a flexible plastic pipe. Thereafter, 1 × 107 HCT116 and SW620 cells in 50–100 μL of 10% Matrigel (Corning, NY, USA)/phosphate buffered saline (PBS) were injected into 8–10-week-old balb/c mice (Male, Orient Bio Inc.). A colonic submucosal injection was gently inserted into the working channel of the endoscope using the modified flexible needle. After injection, tumor formation was confirmed through colonoscopy imaging every week. Colonoscopy was performed using fluorescence endoscopy (Vetcom; Karl Storz).
Real-time fluorescence endoscopy imaging of mouse colon tumor
ICG was purchased from USP (1340009, USP, Rockville, MD, USA), and was dissolved in distilled water with 5% glucose (Sigma-Aldrich Co., St. Louis, MO, USA). All colonoscopies were performed using the animal endoscopic system (Vetcom; Karl Storz). The IMAGE1 H3-ZF1 THREE-CHIP FULL HD Camera system with fluorescent filters (wavelengths 805 nm–835 nm) that enable ICG fluorescence detection in the basic endoscope and the modified D-light PVET source (66100M3) were added. All endoscopy images were recorded as mp4-video files using the AIDA HD control system (Karl Storz) and sequences of endoscopy images were converted into TIF images through snapshots. The endoscopy image fluorescence intensity was analyzed by ImageJ (NIH, Bethesda, MD, USA). Regions of interest (ROIs) were defined around the detected polyp based on the corresponding optical images. Thereafter, ROIs of polyps were compared to adjacent normal mucosa and documented as a tumor-to-normal ratio for each site. After tumor implantation, orthotopic mice were administered ICG via tail vein injection. Thirty minutes later, endoscopy was performed and images were obtained with the animal endoscopic system (Vetcom; Karl Storz). To confirm the intensity of fluorescence detection in ex vivo, molecular imaging was performed using the Xenogen in vivo imaging system (IVIS) spectrum system (Caliper Life Sciences, Inc., Hopkinton, MA, USA).
Ex vivo molecular imaging of CRC patients
Human colon tissue samples (10‒20 mm) were surgically excised from eight patients. Each fresh CRC tissue with adjacent normal colon tissue was evaluated immediately following surgery. Autofluorescence images of the tumor tissues and paired normal colon tissues were obtained prior to ICG incubation. Before incubation, to prevent a possible false-positive effect from probe infiltration of the resection, the tissue was covered with low-melting agarose gel. CRC tissues and paired normal colon tissues were incubated with ICG (5 mg/mL) for 30 min. After incubation, tissues were washed three times with PBS, and fluorescence molecular imaging was performed with the Xenogen IVIS spectrum system (Optix MX3 system; ART Advanced Research Technologies Inc., Montreal, Canada).
Histology and immunohistochemistry
Immunohistochemistry (IHC) was performed using patient colon tissues, which were fixed with 4% paraformaldehyde and embedded in paraffin. Immunostaining was performed using the BenchMark XT automatic immunostaining device (Ventana Medical Systems, Inc., Oro Valley, AR, USA) and OptiView DAB IHC Detection (Ventana Medical Systems, Inc.). Tissue sections (4 μm) were transferred to salinized, charged slides and incubated at room temperature and 65°C. After antigen retrieval for 64 min, the sections were incubated on a fully automated immunostainer with anti-iNOS antibody (Abcam, Cambridge, UK;) for 32 min. Tissue section slides were then counter-stained with 4′,6-diamidino-2-phenylindole (DAPI). All image-staining patterns of the slides were acquired using an OptiView DAB IHC Detection Kit (Ventana Medical Systems, Inc.). Furthermore, all images were acquired using a x20 objective lens.
Immunofluorescence
Tissue cryosections fixed with 4% paraformaldehyde were incubated with iNOS antibody, followed by incubation with anti-rabbit IgG Alexa Fluor TM 488 (Thermo Fisher Scientific). Microscopic images were obtained by using ZEISS Axio Observer (Zeiss, Oberkochen, Germany).
Detection of time-based ICG accumulation in vitro
The tested cell lines (HCT116, SW480, and CCD841) were seeded in 96-well plates at a density of 1 × 104 cells per well in DMEM/ High Glucose medium with 10% FBS. The cells were washed with serum-free media and incubated with ICG (50 μM) in 1% FBS media for various durations (0, 1, 5, 10, 20, 30, and 60 min and 12, 18, and 24 h) at 37°C. After incubation, the cells were washed three times with PBS. To confirm the intensity of ICG accumulation in vitro, fluorescence molecular imaging was performed using the Xenogen IVIS spectrum system (Caliper Life Sciences, Inc.MA, USA).
Western blot analysis
The cell lysates were prepared using radioimmunoprecipitation assay (RIPA) lysis buffer (Thermo Fisher Scientific, Inc.) and protease inhibitor cocktail (GenDEPOT, Barker, TX, USA). Cell lysates were separated using 10% sodium dodecyl sulphate-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to polyvinylidene fluoride (PVDF) membranes. Protein-transferred membranes were blocked with 5% bovine serum albumin (BSA) in Tris-buffered saline Tween 20 (TBS-T) for 1 h at room temperature and incubated with blocking buffer-diluted primary antibodies, including E-cadherin (Abcam; ab76055), iNOS (Abcam; ab178945), c-Met (Abcam; ab51067), VEGFA (Abcam; ab46154), Ki-67 (Abcam; ab16667) and ABCB1 (Abcam; ab170904) antibody (1:1000) at 4°C overnight. After the membranes were washed and incubated with secondary antibodies, immunoreactive protein expression signals were detected using the enhanced chemiluminescence (ECL) substrate (Thermo Fisher Scientific, Inc.) and visualized using Luminograph III (ATTO Corporation, Tokyo, Japan). The expression of anti-β-actin (Sigma-Aldrich Co.) was used as the loading control.
Intracellular nitric oxide assay
The tested cell lines (HCT116, SW480, and CCD841) were seeded in 96-well plates at a density of 1 × 104 cells per well. Quantification of nitric oxide synthase (NOS) activity in a 96-well format was performed using the manufacturer’s instruction. Briefly, the nitric oxide fluorometric probe in working solution was added to each well and incubated at 37°C in the dark for 2 h. Cell lysis buffer was then added to each well without washing before taking the measurement. The fluorescence intensity at excitation/emission wavelengths of 480/530 nm was measured using a CLARIOstar microplate reader (BMG Labtech, Ortenberg, Hessen, Germany).
Artificial intelligence analysis of fluorescence endoscopy images
Sixty images (optical, fluorescence) obtained during endoscopic observations of the AOM/DSS mouse model, together with the deep learning network, were used to categorize four classes of CRC development. First, data preprocessing was performed before inputting it into the network. For data preprocessing, 16 types of augmentation were performed to extract specific features that could improve the accuracy of the network. We aimed for a robust network with high accuracy by adjusting the brightness of the dark image, emphasizing characteristics of the fluorescently marked part, and changing the image shape in various ways. EfficientNet [37] was used as a deep learning network, and a baseline architecture with an input resolution of 224 × 224 was applied. Performance of the network was evaluated using precision, recall, f1-score, and accuracy values. Parameters for network training were 300 epochs, the optimizer was AdamW, and the learning rate was 5.0e-4.
Results
Correlation analysis of fluorescence intensity and molecular targets for ICG fluorescence endoscopy imaging
Colitis-associated colon cancer was induced in 14 mice using AOM/DSS. After a second injection of DSS, all mice were monitored every week with endoscopy, and proliferative lesions of various sizes and phases were observed (S1 Fig in S1 File). At monitoring week 5–6, all mice underwent a colon endoscopy, which revealed multiple tumor polyps that were identified as dysplasia, adenoma, and carcinoma via histological examinations (S2A Fig in S1 File). ICG was used to help visualize and interpret precancerous lesions. After 30 minutes from the ICG injection, the fluorescence signal of most polyps was shown to increase while maintaining a low background, whereas the colon of a normal mouse showed weak signals in both real time fluorescence endoscopy and ex vivo molecular imaging (Fig 1A, S3 Fig in S1 File). To further examine the ICG molecular imaging of tumor growth, the fluorescence intensity value of ROI of 25 tumor lesions and their adjacent normal lesions was quantified based on real time fluorescence endoscopy and ex vivo molecular imaging. Various ICG fluorescence intensity values were detected for most polyps, which significantly increased compared to those of normal lesions (S2B, S2C Fig in S1 File). For the ICG fluorescence intensity values, the average ICG signals of dysplasia, adenoma, and carcinoma were 1.8-fold, 2.8-fold, and 3.1-fold higher than that of the normal lesion, respectively (Table 1). Cases with 1.5-fold higher intensity than the normal lesions were analyzed via real time fluorescence endoscopy; all carcinoma lesions exhibited over 2-fold higher intensity. Furthermore, 90% (9/10 cases) of adenoma and 57% (4/7 cases) of dysplasia were found to have more than 1.5-fold higher fluorescence intensity than that of the adjacent normal lesion via real time fluorescence endoscopy (Table 1, S2 Fig in S1 File).
(A) Representative indocyanine green (ICG) fluorescence endoscopic images and ex-vivo molecular imaging of polyps and adjacent normal lesions from a mouse colon with the corresponding hematoxylin and eosin-stained images (magnification, x200). (B) Immunohistochemistry (IHC) images with VEGFA, Ki-67, CD31, inducible nitric oxide synthase (iNOS), and c-Met for the corresponding lesions (A) (Scale bar: 100 μm). (C) Correlation analysis between endoscopy fluorescence intensity value (AU) of ICG accumulation and the IHC score of Ki-67 and iNOS. (D) Correlation analysis between ex vivo molecular imaging fluorescence intensity values of ICG and the IHC score of Ki-67 and iNOS. (E) Correlation analysis between ICG intensity values and the IHC score of Ki-67 and iNOS based on tumor stages. Data are presented as mean and error (95% confidence interval [CI]). Size of circle, average radiant efficiency from ex vivo molecular images ([p/s/cm2/sr]/[μW/cm2] (×108)).
To further analyze tumor status and ICG fluorescence intensity, we calculated the correlation of fluorescence intensity values, histology, and tumor specific molecules used for molecular imaging of colon cancer. A total of 25 polyps and five samples of normal tissue from 14 AOM/DSS mice were examined for correlation analysis (Fig 1A, S4 Fig in S1 File). IHC of proliferative, cancer specific, vascular targeting, inflammatory molecules, including c-Met, Ki-67, VEGFA, cluster of differentiation 31 (CD31), and iNOS, were scored based on an inspection by a pathologist and relevant references (Fig 1B, S4 Fig in S1 File) [17, 38, 39]. As shown in Fig 1C and 1D, the ICG fluorescence intensity was positively correlated with Ki-67 and iNOS. The correlation coefficient between CD31 and ICG intensity of real time fluorescence endoscopy was 0.447, whereas that with ex vivo molecular imaging was 0.144 (S7 Fig in S1 File). Moreover, results indicate that Ki-67 and iNOS had higher IHC scores for dysplasia than normal tissue, while the highest IHC scores were found for carcinoma, which is consistent with the ICG fluorescence intensity (Fig 1E).
Cellular nitric oxide concentration is related to the preferential accumulation of ICG in tumors
To elucidate the mechanism of preferential retention of ICG in malignant lesions, we examined ICG accumulation in cultured cells depending on incubation time. As shown in Fig 2A, the ICG signal detected in SW480 and HCT116 was significantly stronger than that in the normal CCD841 colon epithelial cell line up until 18 h of treatment. Further, there was no change in cell viability regardless of duration of ICG treatment (S5 Fig in S1 File). The cellular expression of Ki-67, VEGFA, c-Met, and iNOS also increased in SW480 and HCT116, whereas no expression was found in CCD841 (Fig 2B). Cellular nitric oxide is related to the inhibition of drug efflux [40]. After ICG treatment for 30 min, the intracellular concentrations of nitric oxide in SW480 and HCT116 were significantly higher than that in CCD841 (p < 0.0005, Fig 2C). To verify whether the efflux protein ABCB1 (p-gp) was inhibited, change in expression was investigated by immunoblotting, as shown in Fig 2D. Results demonstrate that ABCB1 was down-regulated after treatment with ICG for 30 min in cancer cell lines.
(A) Time-based indocyanine green (ICG) accumulation in a normal cancer cell line (CCD841) and colon cancer cell lines (HCT116 and SW480). A total of 1 × 104 cells/well were treated with 50 μM of ICG for 30 min prior to in vivo imaging system (IVIS) imaging (n = 4) (B) Western blot analyses of VEGFA, Ki-67, inducible nitric oxide synthase (iNOS), and c-Met in SW480, HCT116, and CCD841. (C) Intracellular nitric oxide (NO) assessment in SW480, HCT116, and CCD841 cells. Values are expressed as mean ± standard error of the mean (SEM; *p < 0.05) (D) Western blot analyses of ABCB1 in SW480 and HCT116 with treatment of ICG for 30 min.
Correlation of ICG molecular imaging and iNOS expression in CRC patients
We analyzed the correlation between ICG fluorescence intensity and expression levels of both iNOS and c-Met using 16 samples from eight patients with CRC (Table 2) via ex vivo molecular imaging. After 30 min from ICG treatment, the fluorescence intensity and iNOS expression were measured in tumor and normal tissues (Fig 3A, S6 Fig in S1 File). As shown in Fig 3B, the ICG fluorescence signals of tumor tissues were significantly higher than those of normal tissues (p = 0.0003). Furthermore, the iNOS expression of tumor tissues was increased compared to that of normal tissues and positively correlated with the ICG fluorescence intensity (r = 0.5114, Fig 3C). ICG fluorescence signal and iNOS expression were observed in the same location (Fig 3A). c-Met expression in the tumor was increased as described previously and had a high positive correlation with ICG fluorescence intensity (S7 Fig in S1 File). To assess the effectiveness of ICG for detecting tumor lesions with increased iNOS expression in vivo, orthotopic tumor-bearing mice were evaluated via real time fluorescence endoscopy. Mice with confirmed tumor growth were administered ICG (7.5 mg/kg body weight) via the tail vein. After 30 min, the accumulation of ICG in tumor lesions with minimal autofluorescence background was detected and maintained for 1 h (S8A Fig in S1 File). The expression levels of iNOS and c-Met were detected using IHC staining, which revealed increased expression in tumor with high ICG fluorescence signals (S8B Fig in S1 File).
(A) Representative ex vivo IVIS, hematoxylin and eosin, immunohistochemistry (IHC), and immunofluorescence images of CRC patient tissues with strong (upper panel) and weak (bottom panel) intensity of indocyanine green (ICG). (Scale bar: 50 μm) (B) Average fluorescence intensity of CRC patient specimens treated with ICG (*p < 0.05). (C) Correlation analysis between ICG fluorescence intensity value (average radiant efficiency) and the IHC score of inducible nitric oxide synthase (iNOS) expression in CRC patient tissues.
AI-integrated ICG-fluorescence endoscopy analysis for detection of precancerous lesions in colon cancer development
A total of 60 images (optical, fluorescence) together with histological diagnosis results from the endoscopic observations of the AOM/DSS mouse model were utilized for deep learning network analysis to categorize precancerous lesions into four classes; normal (N), dysplasia (D), adenoma (A), and carcinoma (C; Fig 4A, S4 Fig in S1 File). To prevent overfitting when training the AI on a small number of data, we performed data augmentation including transformations in scale, rotation, luminance/contrast, aspect ratio, and shearing among others (S9 Fig in S1 File). Training was performed with a 40%–60% of the total data set, followed by validation, which was performed with 27%–33% of the total data set (Fig 4B). Classification testing was performed using fluorescence, optical and fluorescence/optical images. Table 3 shows the accuracy of the classification of images into four classes using each image data set; this approach revealed that fluorescence images obtained the highest accuracy (0.750) for class predication of AI. Furthermore, we examined receiver operating characteristic (ROC) curves for three sets of images to assess the performance. As shown in Fig 4C, fluorescence endoscopy images yielded an area under the ROC curve (AUC) of 0.8125, whereas optical and fluorescence/optical images yielded 0.75 and 0.6667, respectively.
(A) Flow chart of image data and study design. (B) Schematic of the training/validation/test process used in this study. (C) Receiver operating curves of the artificial intelligence (AI) classification of 3 groups of images: fluorescence endoscopy, optical endoscopy, and fluorescence/optical images.
Discussion
The preferential accumulation of ICG in tumor has been reported for various tumor types, including colon cancer [18, 21, 26, 41]. However, there are no reports on the potential of ICG-based real time fluorescence endoscopy imaging for detection of precancerous lesions in a clinical setting. In this study, we investigated the visualization ability of ICG endoscopy in colitis-associated colon cancer using an animal model and patient tissue samples. We report a successful trial of AI-integrated diagnosis and classification of precancerous lesions using fluorescence endoscopy. All carcinoma lesions could be detected with an average of 3.10-fold higher intensity; furthermore, 90% of adenoma lesions were detected with more than 1.5-fold higher intensity. The ICG intensity clearly correlated with the expression levels of iNOS, leading to an increase in nitric oxide concentration in tumor cells, which was related with the down-regulation of ABCB1 expression.
Numerous approaches have been employed to detect CRC lesions using fluorescently labeled antibodies or small molecules binding tumor cell [15]. Fluorescence labeled antibodies against EGFR and VEGF showed high specificity for the target tumor; however, these antibodies could not overcome the disadvantages such as longer time to reaction after injection, and immunogenicity [9, 11]. For enhanced ability, small fluorescence labeled peptide targeting c-Met was developed [14] and tested in 15 patients via intravenous injection to observe malignant polyps; however, limitations such as background signal and a low number of cases to confirm safety were reported. In this study, we validated the clinical feasibility of ICG fluorescence endoscopy for tumor detection. ICG has been used in clinics for more than 40 years [42] and the safety of its use in the human body has been confirmed. According to some studies, ICG can be applied to detect colon tumors in a colitis-associated colon cancer animal model [17, 43]. In the present study, the fluorescence intensity of ICG via real time fluorescence endoscopy and ex vivo molecular imaging was quantified, and the correlation coefficient between these values and the tumor growth in 25 polyps from animal models and 16 human tissue samples was calculated. ICG could detect 100% of carcinoma, 90% of adenoma, and 57.1% of dysplasia within 30 min after injection.
Ki-67 labeling is generally used as a marker of proliferation for tumor grading [44, 45] and was found to have a positive correlation with ICG intensity (r = 0.585) in our real time fluorescence endoscopy experiment. The tumor stage of patient tissues with a high percentage of Ki-67 labeling was T4 cancer whereas that of others was T3, as described previously. However, the ICG intensity and Ki-67 labeling were not directly correlated in human patient tumor tissues (r = 0.025). As all patient samples were surgically resected CRC tissues, the ICG intensity could only distinguish between normal and carcinoma tissues rather than cancer grades. The ICG fluorescence intensity was relatively correlated with c-Met expression in animal models and patient tissue samples (r = 0.443 and r = 0.759, respectively). c-Met is considered a suitable biomarker for detection of early stages of cancer, including adenoma [14, 46]. Recently, cyanine dye (Cy5) labeled peptide probe has been tested for fluorescence endoscopy detection in 15 CRC patients, and was reported to successfully detect most adenomatous polyps as well as more than half of hyperplasias [14]. These results are consistent with our correlation data of ICG fluorescence intensity in animal models. The correlation coefficient between ICG fluorescence intensity and vascular targeting molecules, including VEGF (r = 0.250) and CD31 (r = 0.447), was relatively low, which suggests that the tumor selectivity of ICG is related to an increase in vascular permeability rather than an angiogenic effect. Vascular permeability is promoted by inflammatory mediators, such as iNOS [47], which was highly correlated with ICG intensity as per our correlation analysis data. Meanwhile, our data revealed that ICG intensity is highly correlated with iNOS expression in the colitis-associated colon cancer mouse model and tissue samples from CRC patients (r > 0.5). In fact, expression levels of iNOS were demonstrated to be associated with poor survival rates in several cancers [48–50]. In addition, iNOS was shown to be up-regulated in most chronic inflammatory diseases and cancers, including ulcerative colitis and colon cancer [51]. Tumor cellular iNOS contributes to the production of nitric oxide in the presence of nicotinamide adenine dinucleotide phosphate (NADPH) and oxygen. In recent studies, an increase in nitric oxide concentration was shown to mediate drug efflux inhibition by inactivating ATP binding cassette (ABC) transporters, such as ABCB1 (p-gp), ABCC1 (MRP1) and ABCG2 (BCRP) in tumor cells [52–54]. Although previous studies have demonstrated that preferential ICG uptake is due to extravasation and endocytosis [22], the mechanism of tumor selective retention is poorly understood. Based on the in vitro data presented in this present study, concentration of cellular ICG detected in cancer cell lines differed from that of normal colon epithelial cells according to the ICG incubation time. Moreover, both iNOS expression and intracellular nitric oxide concentration were higher in cancer cell lines. We also observed that the expression of ABCB1 was down-regulated in cancer cells after a 30-minute ICG treatment. These results suggest that increased cellular nitric oxide concentration caused differences in ICG efflux in cancer cells compared to that in normal cells, resulting in prolonged retention in the short term (24 h). Indeed, a prior report revealed that the increase in nitric oxide concentration and iNOS expression was associated with a marked reduction in the doxorubicin efflux rate in colon cancer cell lines [40]. Since we have observed differences within 24 hours in in vitro experiments and in colitis-associated colon cancer, it is plausible that alternative mechanisms may exist in various types of cancer. Meanwhile, further research is needed to better understand the role of ICG in imaging inflammatory conditions.
For the first time, we demonstrated that ICG-fluorescence endoscopic images together with AI analysis enable accurate classification of precancerous lesions into four tumor classes. The accuracy of this approach was evaluated by implementation of AI and endoscopic images of colon cancer obtained from the AOM/DSS mouse model which showed the highest AUC (0.8125) by utilizing only fluorescence endoscopy images, while use of optical endoscopy images yielded an AUC of 0.75. Notably, the AUC of the fluorescence/optical endoscopy concurrent analysis was considerably lower (0.6667). These results suggest that ICG positive cells reflect colitis-associated tumor specific characters with intracellular nitric oxide concentration and this feature together with AI training could be utilized to improve clinical diagnostics.
In this study, we demonstrated that ICG together with AI could be employed for carcinoma and adenoma imaging in colitis-associated colon cancer patients. However, this study has some limitations. Notably, a small sample number could limit the universality of the results. Thus, the feasibility of ICG in detecting precancerous lesions in a larger human cohort needs to be investigated.
Conclusion
We showed fluorescence endoscopy images analyzed with AI could detect precancerous lesion of colon cancer with higher accuracy than optical endoscopy images. Furthermore, the positive correlation of ICG intensity and expression of iNOS in patient tumor tissues was demonstrated. We suggested the novel mechanism that intracellular iNOS concentration is related to ICG retention in cancer cells.
Acknowledgments
We thank the Scientific Publications Team at Asan Medical Center for their editorial assistance with the preparation of this manuscript and the Asan Medical Center for their support and instrumentation.
References
- 1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394–424. pmid:30207593.
- 2. Leslie A, Carey FA, Pratt NR, Steele RJ. The colorectal adenoma-carcinoma sequence. Br J Surg. 2002;89(7):845–60. pmid:12081733.
- 3. Itzkowitz SH, Yio X. Inflammation and cancer IV. Colorectal cancer in inflammatory bowel disease: the role of inflammation. Am J Physiol Gastrointest Liver Physiol. 2004;287(1):G7–17. pmid:15194558.
- 4. Kim NH, Jung YS, Jeong WS, Yang HJ, Park SK, Choi K, et al. Miss rate of colorectal neoplastic polyps and risk factors for missed polyps in consecutive colonoscopies. Intest Res. 2017;15(3):411–8. pmid:28670239; PubMed Central PMCID: PMC5478767.
- 5. Atreya R, Goetz M. Molecular imaging in gastroenterology. Nat Rev Gastroenterol Hepatol. 2013;10(12):704–12. pmid:23856892.
- 6. Nguyen QT, Tsien RY. Fluorescence-guided surgery with live molecular navigation—a new cutting edge. Nat Rev Cancer. 2013;13(9):653–62. pmid:23924645; PubMed Central PMCID: PMC4427343.
- 7. Lee JH, Wang TD. Molecular endoscopy for targeted imaging in the digestive tract. Lancet Gastroenterol Hepatol. 2016;1(2):147–55. pmid:28404071.
- 8. Kobayashi H, Ogawa M, Alford R, Choyke PL, Urano Y. New strategies for fluorescent probe design in medical diagnostic imaging. Chem Rev. 2010;110(5):2620–40. pmid:20000749; PubMed Central PMCID: PMC3241938.
- 9. Suo Y, Wu F, Xu P, Shi H, Wang T, Liu H, et al. NIR-II Fluorescence Endoscopy for Targeted Imaging of Colorectal Cancer. Adv Healthc Mater. 2019;8(23):e1900974. pmid:31697035.
- 10. Miyamoto Y, Muguruma N, Fujimoto S, Okada Y, Kida Y, Nakamura F, et al. Epidermal growth factor receptor-targeted molecular imaging of colorectal tumors: Detection and treatment evaluation of tumors in animal models. Cancer Sci. 2019;110(6):1921–30. pmid:30973663; PubMed Central PMCID: PMC6549923.
- 11. Goetz M, Hoetker MS, Diken M, Galle PR, Kiesslich R. In vivo molecular imaging with cetuximab, an anti-EGFR antibody, for prediction of response in xenograft models of human colorectal cancer. Endoscopy. 2013;45(6):469–77. pmid:23580409.
- 12. Kim HI, Kim J, Kim H, Lee H, Yoon YS, Hwang SW, et al. Biomolecular imaging of colorectal tumor lesions using a FITC-labeled scFv-Ckappa fragment antibody. Sci Rep. 2021;11(1):17155. Epub 20210825. pmid:34433835; PubMed Central PMCID: PMC8387423.
- 13. Mucke MM, Bettenworth D, Geyer C, Schwegmann K, Poremba C, Schafers M, et al. Targeting Mucosal Endothelin-A-Receptor Expression by Fluorescence Endoscopy is Feasible to Detect and Characterize Colitis-Associated Cancer in Mice. Inflamm Bowel Dis. 2017;24(1):111–22. pmid:29272493.
- 14. Burggraaf J, Kamerling IMC, Gordon PB, Schrier L, de Kam ML, Kales AJ, et al. Detection of colorectal polyps in humans using an intravenously administered fluorescent peptide targeted against c-Met. Nat Med. 2015;21(8):955–61. Epub 20150713. WOS:000359181000024. pmid:26168295.
- 15. Kashihara T, Muguruma N, Fujimoto S, Miyamoto Y, Sato Y, Takayama T. Recent Advances in Molecular Imaging of Colorectal Tumors. Digestion. 2021;102(1):57–64. pmid:33271567.
- 16. Alford R, Simpson HM, Duberman J, Hill GC, Ogawa M, Regino C, et al. Toxicity of organic fluorophores used in molecular imaging: literature review. Mol Imaging. 2009;8(6):341–54. pmid:20003892.
- 17. Nagahara R, Onda N, Yamashita S, Kojima M, Inohana M, Eguchi A, et al. Fluorescence tumor imaging by i.v. administered indocyanine green in a mouse model of colitis-associated colon cancer. Cancer Sci. 2018;109(5):1638–47. Epub 20180419. pmid:29520973; PubMed Central PMCID: PMC5980401.
- 18. Holt D, Okusanya O, Judy R, Venegas O, Jiang J, DeJesus E, et al. Intraoperative near-infrared imaging can distinguish cancer from normal tissue but not inflammation. PLoS One. 2014;9(7):e103342. Epub 20140729. pmid:25072388; PubMed Central PMCID: PMC4114746.
- 19. Madajewski B, Judy BF, Mouchli A, Kapoor V, Holt D, Wang MD, et al. Intraoperative near-infrared imaging of surgical wounds after tumor resections can detect residual disease. Clin Cancer Res. 2012;18(20):5741–51. Epub 20120829. pmid:22932668; PubMed Central PMCID: PMC3482441.
- 20. Maeda H, Nakamura H, Fang J. The EPR effect for macromolecular drug delivery to solid tumors: Improvement of tumor uptake, lowering of systemic toxicity, and distinct tumor imaging in vivo. Adv Drug Deliv Rev. 2013;65(1):71–9. Epub 20121023. pmid:23088862.
- 21. Kosaka N, Mitsunaga M, Longmire MR, Choyke PL, Kobayashi H. Near infrared fluorescence-guided real-time endoscopic detection of peritoneal ovarian cancer nodules using intravenously injected indocyanine green. Int J Cancer. 2011;129(7):1671–7. Epub 20110618. pmid:21469142; PubMed Central PMCID: PMC3145021.
- 22. Onda N, Kimura M, Yoshida T, Shibutani M. Preferential tumor cellular uptake and retention of indocyanine green for in vivo tumor imaging. Int J Cancer. 2016;139(3):673–82. Epub 20160404. pmid:27006261.
- 23. Gotoh K, Yamada T, Ishikawa O, Takahashi H, Eguchi H, Yano M, et al. A novel image-guided surgery of hepatocellular carcinoma by indocyanine green fluorescence imaging navigation. J Surg Oncol. 2009;100(1):75–9. pmid:19301311.
- 24. Ishizawa T, Masuda K, Urano Y, Kawaguchi Y, Satou S, Kaneko J, et al. Mechanistic background and clinical applications of indocyanine green fluorescence imaging of hepatocellular carcinoma. Ann Surg Oncol. 2014;21(2):440–8. Epub 20131120. pmid:24254203.
- 25. Ntziachristos V, Yodh AG, Schnall M, Chance B. Concurrent MRI and diffuse optical tomography of breast after indocyanine green enhancement. Proc Natl Acad Sci U S A. 2000;97(6):2767–72. pmid:10706610; PubMed Central PMCID: PMC16004.
- 26. Yokoyama J, Fujimaki M, Ohba S, Anzai T, Yoshii R, Ito S, et al. A feasibility study of NIR fluorescent image-guided surgery in head and neck cancer based on the assessment of optimum surgical time as revealed through dynamic imaging. Onco Targets Ther. 2013;6:325–30. Epub 20130408. pmid:23630424; PubMed Central PMCID: PMC3623571.
- 27. Okusanya OT, Holt D, Heitjan D, Deshpande C, Venegas O, Jiang J, et al. Intraoperative near-infrared imaging can identify pulmonary nodules. Ann Thorac Surg. 2014;98(4):1223–30. Epub 20140805. pmid:25106680; PubMed Central PMCID: PMC4185369.
- 28. Tummers QR, Hoogstins CE, Peters AA, de Kroon CD, Trimbos JB, van de Velde CJ, et al. The Value of Intraoperative Near-Infrared Fluorescence Imaging Based on Enhanced Permeability and Retention of Indocyanine Green: Feasibility and False-Positives in Ovarian Cancer. PLoS One. 2015;10(6):e0129766. Epub 20150625. pmid:26110901; PubMed Central PMCID: PMC4482512.
- 29. Liberale G, Vankerckhove S, Caldon MG, Ahmed B, Moreau M, Nakadi IE, et al. Fluorescence Imaging After Indocyanine Green Injection for Detection of Peritoneal Metastases in Patients Undergoing Cytoreductive Surgery for Peritoneal Carcinomatosis From Colorectal Cancer: A Pilot Study. Ann Surg. 2016;264(6):1110–5. pmid:27828822.
- 30. Yanagita T, Hara M, Osaga S, Nakai N, Maeda Y, Shiga K, et al. Efficacy of intraoperative ICG fluorescence imaging evaluation for preventing anastomotic leakage after left-sided colon or rectal cancer surgery: a propensity score-matched analysis. Surg Endosc. 2021;35(5):2373–85. Epub 20210125. pmid:33495878.
- 31. Muguruma N, Miyamoto H, Okahisa T, Takayama T. Endoscopic molecular imaging: status and future perspective. Clin Endosc. 2013;46(6):603–10. Epub 20131119. pmid:24340252; PubMed Central PMCID: PMC3856260.
- 32. Pacal I, Karaboga D, Basturk A, Akay B, Nalbantoglu U. A comprehensive review of deep learning in colon cancer. Comput Biol Med. 2020;126:104003. Epub 20200917. pmid:32987202.
- 33. Davri A, Birbas E, Kanavos T, Ntritsos G, Giannakeas N, Tzallas AT, et al. Deep Learning on Histopathological Images for Colorectal Cancer Diagnosis: A Systematic Review. Diagnostics (Basel). 2022;12(4). Epub 20220329. pmid:35453885; PubMed Central PMCID: PMC9028395.
- 34. Ho C, Zhao Z, Chen XF, Sauer J, Saraf SA, Jialdasani R, et al. A promising deep learning-assistive algorithm for histopathological screening of colorectal cancer. Sci Rep. 2022;12(1):2222. Epub 20220209. pmid:35140318; PubMed Central PMCID: PMC8828883.
- 35. Ozawa T, Ishihara S, Fujishiro M, Kumagai Y, Shichijo S, Tada T. Automated endoscopic detection and classification of colorectal polyps using convolutional neural networks. Therap Adv Gastroenterol. 2020;13:1756284820910659. Epub 20200320. pmid:32231710; PubMed Central PMCID: PMC7092386.
- 36. Girdler B, Moon H, Bae MR, Ryu SS, Bae J, Yu MS. Feasibility of a deep learning-based algorithm for automated detection and classification of nasal polyps and inverted papillomas on nasal endoscopic images. Int Forum Allergy Rhinol. 2021;11(12):1637–46. Epub 20210620. pmid:34148298.
- 37. Tan MX, Le QV. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. Pr Mach Learn Res. 2019;97. WOS:000684034306026.
- 38. Tanaka Y, Ito S, Oshino R, Chen N, Nishio N, Isobe K. Effects of growth arrest and DNA damage-inducible protein 34 (GADD34) on inflammation-induced colon cancer in mice. Br J Cancer. 2015;113(4):669–79. Epub 20150721. pmid:26196182; PubMed Central PMCID: PMC4647681.
- 39. Pan Q, Lou X, Zhang J, Zhu Y, Li F, Shan Q, et al. Genomic variants in mouse model induced by azoxymethane and dextran sodium sulfate improperly mimic human colorectal cancer. Sci Rep. 2017;7(1):25. Epub 20170201. pmid:28154415; PubMed Central PMCID: PMC5453956.
- 40. Riganti C, Miraglia E, Viarisio D, Costamagna C, Pescarmona G, Ghigo D, et al. Nitric oxide reverts the resistance to doxorubicin in human colon cancer cells by inhibiting the drug efflux. Cancer Res. 2005;65(2):516–25. pmid:15695394.
- 41. Martirosyan NL, Cavalcanti DD, Eschbacher JM, Delaney PM, Scheck AC, Abdelwahab MG, et al. Use of in vivo near-infrared laser confocal endomicroscopy with indocyanine green to detect the boundary of infiltrative tumor. J Neurosurg. 2011;115(6):1131–8. Epub 20110916. pmid:21923240.
- 42. Desmettre T, Devoisselle JM, Mordon S. Fluorescence properties and metabolic features of indocyanine green (ICG) as related to angiography. Surv Ophthalmol. 2000;45(1):15–27. pmid:10946079.
- 43. Onda N, Mizutani-Morita R, Yamashita S, Nagahara R, Matsumoto S, Yoshida T, et al. Fluorescence contrast-enhanced proliferative lesion imaging by enema administration of indocyanine green in a rat model of colon carcinogenesis. Oncotarget. 2017;8(52):90278–90. Epub 20171009. pmid:29163827; PubMed Central PMCID: PMC5685748.
- 44. Menon SS, Guruvayoorappan C, Sakthivel KM, Rasmi RR. Ki-67 protein as a tumour proliferation marker. Clin Chim Acta. 2019;491:39–45. Epub 20190114. pmid:30653951.
- 45. Sobecki M, Mrouj K, Colinge J, Gerbe F, Jay P, Krasinska L, et al. Cell-Cycle Regulation Accounts for Variability in Ki-67 Expression Levels. Cancer Res. 2017;77(10):2722–34. Epub 20170310. pmid:28283655.
- 46. Di Renzo MF, Olivero M, Giacomini A, Porte H, Chastre E, Mirossay L, et al. Overexpression and amplification of the met/HGF receptor gene during the progression of colorectal cancer. Clin Cancer Res. 1995;1(2):147–54. pmid:9815967.
- 47. Maeda H. Vascular permeability in cancer and infection as related to macromolecular drug delivery, with emphasis on the EPR effect for tumor-selective drug targeting. Proc Jpn Acad Ser B Phys Biol Sci. 2012;88(3):53–71. pmid:22450535; PubMed Central PMCID: PMC3365245.
- 48. Ekmekcioglu S, Ellerhorst JA, Prieto VG, Johnson MM, Broemeling LD, Grimm EA. Tumor iNOS predicts poor survival for stage III melanoma patients. Int J Cancer. 2006;119(4):861–6. pmid:16557582.
- 49. Lagares-Garcia JA, Moore RA, Collier B, Heggere M, Diaz F, Qian F. Nitric oxide synthase as a marker in colorectal carcinoma. Am Surg. 2001;67(7):709–13. pmid:11450795.
- 50. Wang J, He P, Gaida M, Yang S, Schetter AJ, Gaedcke J, et al. Inducible nitric oxide synthase enhances disease aggressiveness in pancreatic cancer. Oncotarget. 2016;7(33):52993–3004. pmid:27367029; PubMed Central PMCID: PMC5288163.
- 51. Ohshima H, Tazawa H, Sylla BS, Sawa T. Prevention of human cancer by modulation of chronic inflammatory processes. Mutat Res. 2005;591(1–2):110–22. Epub 20050803. pmid:16083916.
- 52. El-Awady R, Saleh E, Hashim A, Soliman N, Dallah A, Elrasheed A, et al. The Role of Eukaryotic and Prokaryotic ABC Transporter Family in Failure of Chemotherapy. Front Pharmacol. 2016;7:535. Epub 20170110. pmid:28119610; PubMed Central PMCID: PMC5223437.
- 53. Sinha BK, Bortner CD, Mason RP, Cannon RE. Nitric oxide reverses drug resistance by inhibiting ATPase activity of p-glycoprotein in human multi-drug resistant cancer cells. Biochim Biophys Acta Gen Subj. 2018;1862(12):2806–14. Epub 20180901. pmid:30251669; PubMed Central PMCID: PMC6195836.
- 54. Kim M, Park SC, Lee DY. Glycyrrhizin as a Nitric Oxide Regulator in Cancer Chemotherapy. Cancers (Basel). 2021;13(22). Epub 20211117. pmid:34830916; PubMed Central PMCID: PMC8616433.