IMPROVEMENT OF GASTRIC CANCER DETECTION USING ARTIFICIAL INTELLIGENCE WITH LINKED COLOR IMAGING

Background and Aims: To date, whole gastrointestinal capsule endoscopy from the esophagus to the colorectum has not been fully accomplished. We have developed the guiding device using neodymium magnet (L 139 mm X W 139 mm X H 30 mm, weight: 2.3 kg) by which a colon capsule endoscope (PillCam^TMCOLON2) can be manipulated 10 cm apart with 6.3-mN propulsion force exerted by 17-mT magnetic flux density in a single hand. We aim to determine the diagnostic yields of this magnetically-aided capsule endoscopy (MACE) for the esophageal, gastric and colorectal tumors in a randomized controlled trial.
Methods: Patients with superficial esophageal carcinomas (EC), gastric tumors (GT), and colorectal tumors (CRT) were randomly allocated to two groups: one group with a same-sized non-magnetic sham device (S group) and the other with the magnetic device (M group) (1:1; stratified by age, gender, and tumor location). CE was performed mainly in the stomach after ingesting polyethylene glycol-electrolyte lavage solution plus ascorbic acid and mineral water. Optical colonoscopy was performed after excreting CE or around 5 pm without excretion on the same examination day. The primary outcomes were detection rates for the documented tumors and unknown gastrointestinal lesions, based on esophagogastroduodenoscopic and colonoscopic findings. The secondary outcome was observation rates of each gastric part (jRCTs042180150).
Results: We randomly allocated 92 patients (74 men and 18 women, median age: 67, 20-78 y) to the S group (n= 45; 5 ECs, 22 GTs: 7 submucosal and 12 intramucosal carcinomas, and 3 dysplasias, 18 CRTs: 2 advanced carcinomas, 1 submucosal carcinoma, and 15 dysplasias) and the M group (n=47; 4 ECs, 25 GTs: 2 advanced, 1 submucosal, and 16 intramucosal carcinomas, 4 dysplasias, 1 MALToma, and 1 hamartoma, and 18 CRTs: 3 intramucosal carcinomas, 13 dysplasias, 1 hamartoma, and 1 serrated lesion). Examination time and excretion rates were 308 (74-533) m and 315 (84-867) m (P=0.0670), and 87% and 85% (P>0.9999) in the S and M groups, respectively. Detection rates for GTs in the S and M groups were 95% and 100% (P=0.4565), while those for unknown gastric lesions such as polyps and xanthomas were 73％ (47/64) and 100％(81/81, P＜0.0001), even though median numbers of recoding images in the stomach were 98,962 (3,054 - 233,406) and 94,301 (15,213 - 335,827, P=0.7821), respectively. Detection rates for CRTs were 100% and 94% (P>0.9999), and those for unknown colorectal lesions were 75％ and 68％ (P=0.3787); those for ECs were 60% and 75% (P>0.9999) in the S and M groups, respectively. Gastric observation rates were 85% and 97% (P< 0.0001) in the S and M groups, respectively, in which those of the fornix, lower body, and angle were significantly distinct.
Conclusions: This trial provides evidence supporting that MACE is feasible to undergo gastrointestinal tumor screening.

Background: Magnetically assisted capsule endoscopy (MACE) is emerging as an option to complement and replace the upper endoscopy. Three-dimension (3D) MACE system can reconstruct 3D images and improve image quality compared to conventional MACE. We decided to identify the performance and safety of novel 3D MACE and hand-held controller for upper gastrointestinal and small bowel examination.
Method: This study was organized as a prospective, single-center, sequential examination, and non-inferiority trial (KCT0007114). It was conducted on patients who visited Dongguk University Ilsan Hospital for upper endoscopy between April and June 2022. First, 3D MACE was performed for upper gastrointestinal examination. Then two hours later, a sedative upper endoscopy was performed to confirm the accuracy and safety of 3D MACE. Small bowel examination was performed simultaneously using 3D MACE. The primary outcome was to identify of major gastric structures: Esophagogastric junction, cardia/fundus, body, angle, antrum, and pylorus. The secondary outcomes included identification of major upper gastrointestinal structures (add esophagus and bulb from primary outcome), completion of small bowel examination, 3D image reconstruction of gastric lesion, and confirmation of 3D MACE safety.
Results: Upper gastrointestinal examination time was 14.84 ± 3.02 mins in the 3D MACE group and 5.22 ± 2.39 mins in the upper endoscopy. The success rate of identification of the six major gastric structures was 98.6% in 3D MACE and 100% in upper endoscopy. The success rate of identification of the eight major upper gastrointestinal structures was 99.1% in 3D MACE and 100% in upper endoscopy, showing similar performance (Figure 1). 3D reconstructed images were acquired for all lesions observed in 3D MACE using a reading viewer (Figure 2). A complete small bowel examination was confirmed in 52 of 55 patients (94.5%), and the mean small bowel transit time was 260.34 ± 101.14 minutes. For overall patient satisfaction based on a maximum of 10 points (the higher the better), 3D MACE was scored 9.55 ± 0.79 and upper endoscopy was scored 7.75 ± 2.34 (p<0.0001). No significant adverse event and capsule retention occurred in the 3D MACE system.
Conclusion: Novel 3D MACE with hand-held magnet controller is a non-inferior option for upper gastrointestinal examination compared to upper endoscopy. In addition, small bowel examination was possible simultaneously with upper gastrointestinal examination with 3D MACE.

BACKGROUND & AIMS: During wireless capsule endoscopy, real-time localization of the capsule endoscope in the alimentary tract depends on visual inspection of endoscopic images, which is a laborious task. To date, computer-aided diagnosis (CAD) systems to localize the capsule endoscope have not been realized in clinical practice. In this study, a CAD system for localization of the capsule endoscope based on endoscopic images was developed and its diagnostic performance evaluated.
METHODS: A total of 199 capsule endoscopic videos consecutively recorded from October 2010 through August 2021 in one institution were assigned as training data (n=172), validation data (n=7), and testing data (n=20). Still images (stomach 22640, small intestine 24140, and colon 24842) were randomly extracted from training data, and preprocessed using data augmentation to improve data versatility. Feature extraction was performed by a deep neural network (SwinTransformer) to construct a CAD system that discriminates images of the stomach, small intestine (including duodenum) and colon. The reference standard for localization is based on visual inspection by physicians. The CAD system was trained to classify the location of endoscopy images, and the corresponding accuracy was evaluated using still images extracted from validation videos. Random errors of image classification were further removed by an isolation tree, an unsupervised tree-based algorithm for anomaly detection. The performance of the final CAD system was evaluated using 19502 video clips (stomach 1907, small intestine 15494, large intestine 2101) from 20 testing videos. Each video clip consisted of 200 consecutive frame images, and neighboring clips have an overlap of 100 images. The average classification result of all frame images in a video clip was calculated for final determination of the location. Diagnostic performance is described as value [95% confidence interval].
RESULTS: The classification accuracies of video clips were 96.4% [95.4-97.2], 99.3% [99.2-99.4], and 99.8% [99.5-99.9] for stomach, small intestine, and colon, respectively, and the overall accuracy is 99.1% [98.9-99.2]. A detailed review of misdiagnosed video clips revealed that the stomach was misidentified as small intestine in 3.6% [2.8-4.6]; small intestine was misidentified as the stomach in 0.34% [0.26-0.45] and large intestine in 0.39% [0.30-0.50]; large intestine mucosa was misidentified as small intestine in 0.19% [0.05-0.49]. The misidentified video clips tended to have more residue and mucus in the lumen.
CONCLUSIONS: The CAD system developed here provides information to localize the capsule endoscope with high accuracy based on endoscopic images, warranting clinical implementation.

Background: The introduction of motorized spiral enteroscopy (mSe) into clinical practice holds diagnostic and therapeutic potential for small bowel investigation. This systematic review with meta-analysis aims to evaluate the performance of this new modality in diagnosing and treating small bowel lesions.

Material and methods: A systematic search using Medline and Cochrane databases for relevant studies were performed through September 2022. The primary outcome was diagnostic success, defined as the identification of a lesion relative to the indication. Secondary outcomes included successful therapeutic manipulations, total enteroscopy rate (examination from the duodenojejunal flexion to the cecum), technical success (passage from the ligament of Treitz or ileocecal valve for anterograde and retrograde approach, respectively) and complication rates. We performed meta-analyses using a random effects model and results were reported as percentages with 95% Confidence Intervals (95%CIs).

Results: From 2016 to 2022 nine studies (959 patients; 42% females; mean age>45 years; and most commonly investigated for mid GI bleeding/anemia) were considered eligible and included in the analysis. The diagnostic success rate of mSE was 78% (95%CI: 72-84). Considering secondary outcomes, total enteroscopy was attempted in 460 cases, and completed with a rate of 51% (95%CI: 30-72), whereas therapeutic interventions were successful in 98% (95%CI: 96-100) of cases where attempted. Technical success rates were 96% (95%CI:94-97) for anterograde and 97% (95%CI: 94-100) for retrograde approaches, respectively. Finally, the incidence of complications was 17% (95%CI: 13-21), albeit the vast majority included minor adverse events [16% (95%CI: 11-20) vs major= 1% (95%CI: 0-1)].

Conclusion: Motorized spiral enteroscopy provides high rates of diagnostic and therapeutic success with a low prevalence of severe adverse events.

Background and aims: Recent prospective study have shown that novel motorized spiral enteroscopy (NMSE) enables deeper and total small bowel evaluation compared to single-balloon enteroscopy (SBE) in suspected Crohn’s disease (CD) when analyzed per procedure. However, no randomized controlled study have compared bidirectional NMSE with bidirectional SBE in suspected CD.
Methods: Patients with suspected CD requiring small bowel enteroscopy were randomly assigned to either SBE or NMSE between May 2022 to September 2022 in a high volume tertiary center. Bidirectional enteroscopy was done if intended lesion can not be reached on uni-directional study. Comparison was made with regard to technical success (ability to reach lesion), diagnostic yield, depth of maximal insertion (DMI), procedure time and total enteroscopy rates. Depth: time ratio was calculated to avoid confounding for the location of lesion.
Results: Among 125 suspected CD patients (28% female, 18-65 years, median 41 years), 62 and 63 underwent NMSE and SBE respectively (figure 1). The overall technical success was not significantly different between the two (98.4 %: NMSE; 90.5 %: SBE; p=0.11). However there was a trend towards higher technical success with NMSE (94.1%) than SBE (66.7%)(p=0.08) for proximal ileum lesions (Figure 1 and 2A) . There was no significant difference in the diagnostic yield ( 95.2%: NMSE; 87.3%: SBE, p=0.2) (Figure 2B). DMI was higher with NMSE [DMI (cm): median (range): NMSE: 500(80-600), SBE: 180(50-600),p <0.0001] (Figure 2C). Although procedure time was similar [duration (min): median (range): NMSE: 40(7-145),SBE: 45 (13-260),p=0.32], depth: time ratio was significantly higher for NMSE compared to SBE [Depth:time (cm/min): median (range): NMSE: 11.5 (2.67-27.5), SBE: 3.33(1-13.64), p <0.0001] (Figure 2D). Of the 27 patients where total enteroscopy was attempted with NMSE, success was achieved in 77.8% (21/27) [11.1% (1/9) with SBE, p=0.0007]. All adverse events were mild.
Conclusions: NMSE and SBE have comparable technical success and diagnostic yield for small bowel evaluation in suspected CD. NMSE scores over SBE with regards to deeper small bowel evaluation with complete small bowel coverage and higher depth of insertion in shorter time (NCT05363930).

Background: Miss rate of gastric cancer is reported to be approximately 4-26% during esophagogastroduodenoscopy. Linked color imaging (LCI) in LASER endoscopes (Fujifilm, Tokyo, Japan) has been reported to improve gastric cancer detection. On the other hand, some automatic detection systems of artificial intelligence (AI) have been available for the detection of colorectal tumors under white light observation (WLI). However, AI detection system for gastric cancer was not available in daily clinical situation. In this study, we compared the ability of AI and endoscopists on gastric cancer detection by using videos of gastric caners taken by WLI and LCI.
Methods: In our study, a training dataset was composed of 8859 images using non-magnifying WLI and LCI in 405 patients with 462 gastric cancers diagnosed in 2017-2020. A computer-aided detection (CADe) system based on pretrained YOLOv5 was finetuned by the training dataset. In addition, a test set consisted of WLI and LCI videos was obtained from 32 consecutive patients with gastric cancers diagnosed in 2017-2020 (average tumor diameter: 20 mm, macroscopic type: elevated type 14 lesions, depressed type 16 lesions, flat type 2 lesions, histology: well-differentiated 21 lesions, moderately differentiated 5 lesions, poorly differentiated 5 lesions, adenoma 1 lesion, depth of invasion: intramucosa 28 lesions and submucosa 10 lesions). All 32 videos were evaluated by the CADe, two experts with over 15 years’ endoscopy experiences, and two non-expert endoscopists. The comparison between CADe and the endoscopist was also performed.
Results: The sensitivities of CADe were 78.1% and 90.6% for WLI and LCI, respectively with a significant difference (p<0.001). The specificities of CADe were 93.6% and 92.8% for WLI and LCI, respectively with no significant difference (p=0.128). The sensitivities of non-expert endoscopists were 53.1% and 70.3% for WLI and LCI, respectively. Those of expert endoscopists were 73.4% and 92.2% for WLI and LCI, respectively. The sensitivities of CADe for WLI and LCI were significantly higher than those of non-experts for WLI and LCI (p=0.035 and p=0.047, respectively). In contrast, the sensitivities for WLI and LCI had no significant difference between CADe and experts (p=0.134 and p=0.668, respectively).
Limitation: This retrospective study may have different results from those of real endoscopy because of a selection bias in validation.
Conclusions: LCI with CADe demonstrates a gastric cancer detectability better than that of WLI with CADe, but similar to that of expert endoscopists.