Real-Time Use of Artificial Intelligence in Identification of Diminutive Polyps During Colonoscopy

By Y. Mori, S. Kudo, M. Misawa, Yutaka Saito, H. Ikematsu, K. Hotta, K. Ohtsuka, F. Urushibara, S. Kataoka, Y. Ogawa and 12 more. Annals of Internal Medicine, 2018

415 citations recorded by the source corpus.

Colonoscopy is considered the gold standard for decreasing colorectal cancer incidence and mortality (1). To maximize the benefits and cost-effectiveness of colonoscopy, resection of all colorectal polyps except diminutive (5 mm), nonneoplastic rectosigmoid polyps is recommended (2). To allow a diagnose-and-leave strategy for these nonneoplastic (or hyperplastic) polyps, the American Society of Gastrointestinal Endoscopy in its second Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI-2) declaration proposed a negative predictive value (NPV) threshold of 90% or greater for real-time optical diagnosis using advanced endoscopic imaging for detecting diminutive, neoplastic rectosigmoid lesions (such as adenomas) (2). A meta-analysis conducted by the society (3) showed that optical diagnosis of diminutive polyps with narrow-band imaging (NBI [Olympus]; application of narrow-spectrum optical filters at the push of the endoscope's button) in expert hands exceeded the 90% NPV threshold at 93% (95% CI, 91% to 96%), although nonexperts do not perform as well (NPV, 87% [CI, 83% to 91%]). Therefore, a major problem with optical diagnosis is its standardization. To overcome this limitation, computer-aided diagnosis (CAD) powered by artificial intelligence is attracting attention as an option to improve optical diagnosis (4, 5). Several CAD systems designed for use with magnifying NBI showed 90% or greater NPV for diagnosing diminutive adenomas (610). Most of these studies, however, were conducted in experimental, retrospective settings, in which accuracy tends to be greater than in the clinical setting owing to selection bias. Hence, the role of CAD in a colonoscopy practice has not been established. This prospective study assessed the feasibility and accuracy of CAD in assessing diminutive polyps during colonoscopy in a large group of patients. We primarily aimed to clarify whether CAD designed for use with endocytoscopes (520 ultramagnifying endoscopes [Olympus]) could exceed the 90% or greater NPV threshold for diagnosing diminutive rectosigmoid adenomas sufficiently to adopt the diagnose-and-leave strategy of PIVI-2. Methods Study Design and Sites This study was a single-group, open-label, prospective study. Patient enrollment and colonoscopy procedures were conducted at a single center (Showa University Northern Yokohama Hospital). The CAD system was trained by using endoscopic data acquired from 5 centers (National Cancer Center Hospital, National Cancer Center Hospital East, Shizuoka Cancer Center, Tokyo Medical and Dental University, and Showa University Northern Yokohama Hospital). Patients Consecutive patients scheduled to undergo routine colonoscopy were recruited to the study between June and December 2017. Only those aged 18 years or older, not receiving anticoagulant therapy, and able to give informed consent were eligible to participate. Exclusion criteria included a history of inflammatory bowel disease, chemotherapy, or radiation therapy for colorectal cancer. Endoscopy Procedures The study participants underwent colonoscopy with an endocytoscope under conscious sedation (diazepam, 5 to 10 mg intravenously). Endoscopists were not required to meet any qualification level. While withdrawing the colonoscope from the cecum, endoscopists predicted the pathology of the detected polyps with on-site use of CAD (EB-01 prototype [Cybernet Systems]), whose output (neoplastic or nonneoplastic) was displayed in real time on the endoscopy monitor. The endoscopists used CAD first with the NBI mode (CAD-NBI) followed by the stained mode (CAD-stained) and were encouraged to capture 10 or more endocytoscopic images of each polyp (for details, see the sections Endocytoscope and Real-Time Working Artificial Intelligence System). The polypsserving as subjects of the analysiswere then resected and evaluated pathologically. The ability of CAD to differentiate neoplastic from nonneoplastic polyps was assessed, with pathologic findings as the gold standard. Registration of Case Report Forms Immediately after each colonoscopy, endoscopists completed case report forms by inputting the size, location, morphology, and resection technique for each detected polyp, together with the insertion time and any complications. The pathologic prediction output by CAD for each polyp was automatically recorded in a comma-separated value file in the computer connected to the endoscopy unit. Two research assistants were engaged in inputting these data into the database. One assistant input data from the endoscopypathology case reports, while the other, who was blinded to these reports, input the data from the comma-separated value file. Endocytoscope We used a prototype endocytoscope (CF-Y-0058; the same model is commercially available as CF-H290ECI [Olympus]) to acquire images for CAD analysis. This endoscope has a contact light microscopy system (magnification, 520; focusing depth, 35 m; field of view, 570500 m) integrated into the distal tip of a colonoscope, enabling in vivo microvascular evaluation with the NBI mode and cellular visualization after staining with 1.0% methylene blue (stained mode) simply by pulling a hand-operated lever (Appendix Figure 1). The endocytoscope also functions as a normal, high-definition endoscope that allows the same insertion, observation, and treatment as a traditional colonoscope. Appendix Figure 1. Endocytoscopic images. A diminutive adenoma (A to C) and hyperplastic polyp (D to F) are seen in the rectum on white light (A and D) and endocytoscopic images (magnification, 520) in NBI (B and E) and stained mode (C and F). Endocytoscopic images of the adenoma reveal networking microvessels in NBI mode and slit-like lumens (arrowheads) with slightly swollen nuclei (arrows) in stained mode. In contrast, endocytoscopic images of the hyperplastic polyp reveal obscure microvessels in NBI mode and serrated lumens (arrowheads) and small intracellular granules (arrows) in stained mode. NBI = narrow-band imaging. Real-Time Working Artificial Intelligence System The algorithm of the CAD model (Figure 1) was developed on the basis of our previous study (8, 9). The computer with the installed CAD system was connected directly to an endoscopy unit (Evis Lucera Elite CV 290 [Olympus]), allowing fully automated, real-time diagnosis (Supplement Video). The CAD algorithm has 3 steps: feature extraction, classification, and pathologic prediction. Figure 1. Real-time CAD for use with an endocytoscope (CF-Y0058-I/CF-H290ECI [Olympus]) and its algorithm. The computer on which CAD is installed is connected directly to an endoscopy unit (Evis Lucera Elite CV 290 [Olympus]), allowing fully automated diagnosis. Endoscopists can initiate CAD output simply by pushing the capture button on the endoscope. The algorithm for CAD in both NBI and stained modes comprises 3 steps: texture analysis, classification, and pathologic prediction. These analyses take 0.4 seconds, immediately outputting the pathologic prediction of the target polyp as either neoplastic or nonneoplastic, along with the probability of the diagnosis (0% to 100%). CAD= computer-aided diagnosis; NBI= narrow-band imaging. During feature extraction, 312 variables were drawn from the acquired endocytoscopic image by analyzing textures that were characterized by contrast differences of each adjacent pixel. Arrangements of vessels and cellular images thus were quantified indirectly. After extraction, a support vector machine classified the imaged polyp as neoplastic or nonneoplastic on the basis of the 312 extracted features. Our support vector machine, a machine-learning method for optimal separation of complex objects (11), was trained by using endocytoscopic images provided by the aforementioned 5 academic centers. During the study period, we updated the CAD system 5 times by periodically adding training images for machine learningstarting with 28152 images (1 June 2017) and ending with 61925 (5 December 2017). The predicted pathologic results were finally displayed as neoplastic or nonneoplastic, along with the probability of each prediction (0% to 100%) as estimated by the support vector machine. Endoscopists were encouraged to capture several images of each polyp, so the final CAD result was based on majority rule (that is, if 6 outputs classified the polyp as neoplastic and 2 as nonneoplastic, the final diagnosis was neoplastic). If the quality of the captured image was not appropriate for automated diagnosis, CAD automatically recognized it as nonanalyzable, showing Not a good sample on the monitor. Outcome Measures The primary end point was whether CAD-stained analysis produced a 90% or greater NPV for diagnosing diminutive rectosigmoid adenomas. We selected the performance of CAD-stained analysis as the primary outcome measure, because a retrospective study showed that its diagnostic ability reached the PIVI-2 threshold (9), whereas we had no substantial corresponding data for CAD-NBI (8). Secondary outcome measures included the ability of CAD-NBI to distinguish diminutive rectosigmoid polyps and the technical success rate of performing CAD. We also recorded the cecal intubation rate, insertion time, and complications related to insertion to verify the basic function of the endocytoscope. This study had 2 additional co-primary end points: whether trainees' image acquisition rates were higher when they performed endocytoscopy with NBI versus magnified endoscopy with NBI for diminutive polyps and whether CAD-stained analysis had a positive predictive value greater than 90% for diagnosing invasive colorectal lesions 20 mm or larger. We plan to report these end points in separate publications. Pathology Pathologists blinded to the optical diagnosis classified the specimens according to the World Health Organization Classification of Tumours, 4th edition (12). We excluded sessile serrated adenomas and polyps (SSA/Ps) and pathologically heterogeneous lesio

Subfields: Artificial Intelligence, Oncology, Radiology, Nuclear Medicine and Imaging

View Paper: https://www.semanticscholar.org/paper/99f2d8ba2ad1224ff08b31583b007580a2c48133

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Oncology

Douglas Rex

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

Key challenges for delivering clinical impact with artificial intelligence (2019, background): “The limited number of prospective studies to date include diabetic retinopathy grading [48 – 50], detection of breast cancer metastases in sentinel lymph node biopsies [51, 52], wrist fracture detection [53], colonic polyp detection [28, 54], and detection of congenital cataracts [55].”
Performance of artificial intelligence for colonoscopy regarding adenoma and polyp detection: a meta-analysis. (2020, methodology): “Initial studies demonstrated favorable results on detection 10–12 , but were generally underpowered to assess the relationship between increased detection and lesion characteristics, such as polyp size, morphology, location or histology.”
Role of Artificial Intelligence Applications in Real-Life Clinical Practice: Systematic Review (2020, background): “Seven studies were aimed at clinician workflow and efficiency [26,27,44,51,54,66,73].”
High-performance medicine: the convergence of human and artificial intelligence (2019, methodology): “The first prospective clinical validation of AI was performed in 325 patients who collectively had 466 tiny polyps, with an accuracy of 94% and negative predictive value of 96% during real-time, routine colonoscopy 36,64 .”
Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study (2019, methodology): “In the future, this CADe system may be combined with a CADx system to support a detect, diagnose and disregard 43 strategy to avoid excessive workload.”