Multi-step validation of a Post-Endoscopic Retrograde Cholangiopancreatography Pancreatitis prediction system integrating multi-modal data: A multi-center study.

BACKGROUND AND STUDY AIMS: The impact of various categories of information on the prediction of Post Endoscopic Retrograde Cholangiopancreatography Pancreatitis (PEP) remains uncertain. We aimed to comprehensively investigate the risk factors associated with PEP by constructing and validating a model incorporating multi-modal data through multiple steps. PATIENTS AND METHODS: A total of 1,916 cases underwent ERCP were retrospectively collected from multiple centers for model construction. Through literature research, 49 electronic health record (EHR) features and one image feature related to PEP were identified. The EHR features were categorized into baseline, diagnosis, technique, and prevent strategies, covering pre-ERCP, intra-ERCP, and peri-ERCP phases. We first incrementally constructed models 1-4 incorporating these four feature categories, then added the image feature into models 1-4 and developed models 5-8. All models underwent testing and comparison using both internal and external test sets. Once the optimal model was selected, we conducted comparison among multiple machine learning algorithms. RESULTS: Compared with model 2 incorporating baseline and diagnosis features, adding technique and prevent strategies (model 4) greatly improved the sensitivity (63.89% vs 83.33%, p<0.05) and specificity (75.00% vs 85.92%, p<0.001). Similar tendency was observed in internal and external tests. In model 4, the top three features ranked by weight were previous pancreatitis, NSAIDS, and difficult cannulation. The image-based feature has the highest weight in model 5-8. Lastly, model 8 employed Random Forest algorithm showed the best performance. CONCLUSIONS: We firstly developed a multi-modal prediction model for identifying PEP with clinical-acceptable performance. The image and technique features are crucial for PEP prediction.

Validation of artificial intelligence-based bowel preparation assessment in screening colonoscopy (with video).

BACKGROUND AND AIMS: Accurate bowel preparation assessment is essential for determining colonoscopy screening intervals. Patients with suboptimal bowel preparation are at a high risk of missing >5 mm adenomas and should undergo an early repeat colonoscopy. In this study, we used artificial intelligence (AI) to evaluate bowel preparation and validated the ability of the system to accurately identify patients who are at high risk of having >5 mm adenomas missed due to inadequate bowel preparation. METHODS: This prospective, single-center, observational study was conducted at the Eighth Affiliated Hospital, Sun Yat-sen University, from October 8, 2021, to November 9, 2022. Eligible patients who underwent screening colonoscopy were consecutively enrolled. The AI assessed bowel preparation using the e-Boston Bowel Preparation Scale (e-BBPS) while endoscopists made evaluations using BBPS. If both BBPS and e-BBPS deemed preparation adequate, the patient immediately underwent a second colonoscopy; otherwise, the patient underwent bowel re-cleansing before the second colonoscopy. RESULTS: Among the 393 patients, 72 adenomas >5 mm in size were detected; 27 adenomas >5 mm in size were missed. In unqualified-AI patients, the >5 mm adenoma miss rate (AMR) was significantly higher than in qualified-AI patients (35.71% vs 13.19% [P = .0056]; odds ratio [OR], .2734 [95% CI, .1139-.6565]), as were the AMR (50.89% vs 20.79% [P < .001]; OR, .2532 [95% CI, .1583-.4052]) and >5 mm polyp miss rate (35.82% vs 19.48% [P = .0152]; OR, .4335 [95% CI, .2288-.8213]). CONCLUSIONS: This study confirmed that patients classified as inadequate by AI exhibited an unacceptable >5 mm AMR, providing key evidence for implementing AI in guiding bowel re-cleansing and potentially standardizing future colonoscopy screening. (Clinical trial registration number: NCT05145712.).

Effect of artificial intelligence on novice-performed colonoscopy: a multicenter randomized controlled tandem study.

BACKGROUND AND AIMS: The efficacy and safety of colonoscopy performed by artificial intelligence (AI)-assisted novices remain unknown. The aim of this study was to compare the lesion detection capability of novices, AI-assisted novices, and experts. METHODS: This multicenter, randomized, noninferiority tandem study was conducted across 3 hospitals in China from May 1, 2022, to November 11, 2022. Eligible patients were randomized into 1 of 3 groups: the CN group (control novice group, withdrawal performed by a novice independently), the AN group (AI-assisted novice group, withdrawal performed by a novice with AI assistance), or the CE group (control expert group, withdrawal performed by an expert independently). Participants underwent a repeat colonoscopy conducted by an AI-assisted expert to evaluate the lesion miss rate and ensure lesion detection. The primary outcome was the adenoma miss rate (AMR). RESULTS: A total of 685 eligible patients were analyzed: 229 in the CN group, 227 in the AN group, and 229 in the CE group. Both AMR and polyp miss rate were lower in the AN group than in the CN group (18.82% vs 43.69% [P < .001] and 21.23% vs 35.38% [P < .001], respectively). The noninferiority margin was met between the AN and CE groups of both AMR and polyp miss rate (18.82% vs 26.97% [P = .202] and 21.23% vs 24.10% [P < .249]). CONCLUSIONS: AI-assisted colonoscopy lowered the AMR of novices, making them noninferior to experts. The withdrawal technique of new endoscopists can be enhanced by AI-assisted colonoscopy. (Clinical trial registration number: NCT05323279.).

Assessment of the role of false-positive alerts in computer-aided polyp detection for assistance capabilities.

BACKGROUND AND AIM: False positives (FPs) pose a significant challenge in the application of artificial intelligence (AI) for polyp detection during colonoscopy. The study aimed to quantitatively evaluate the impact of computer-aided polyp detection (CADe) systems' FPs on endoscopists. METHODS: The model's FPs were categorized into four gradients: 0-5, 5-10, 10-15, and 15-20 FPs per minute (FPPM). Fifty-six colonoscopy videos were collected for a crossover study involving 10 endoscopists. Polyp missed rate (PMR) was set as primary outcome. Subsequently, to further verify the impact of FPPM on the assistance capability of AI in clinical environments, a secondary analysis was conducted on a prospective randomized controlled trial (RCT) from Renmin Hospital of Wuhan University in China from July 1 to October 15, 2020, with the adenoma detection rate (ADR) as primary outcome. RESULTS: Compared with routine group, CADe reduced PMR when FPPM was less than 5. However, with the continuous increase of FPPM, the beneficial effect of CADe gradually weakens. For secondary analysis of RCT, a total of 956 patients were enrolled. In AI-assisted group, ADR is higher when FPPM ≤ 5 compared with FPPM > 5 (CADe group: 27.78% vs 11.90%; P = 0.014; odds ratio [OR], 0.351; 95% confidence interval [CI], 0.152-0.812; COMBO group: 38.40% vs 23.46%, P = 0.029; OR, 0.427; 95% CI, 0.199-0.916). After AI intervention, ADR increased when FPPM ≤ 5 (27.78% vs 14.76%; P = 0.001; OR, 0.399; 95% CI, 0.231-0.690), but no statistically significant difference was found when FPPM > 5 (11.90% vs 14.76%, P = 0.788; OR, 1.111; 95% CI, 0.514-2.403). CONCLUSION: The level of FPs of CADe does affect its effectiveness as an aid to endoscopists, with its best effect when FPPM is less than 5.

The effect of incorporating domain knowledge with deep learning in identifying benign and malignant gastric whitish lesions: A retrospective study.

BACKGROUND AND AIM: Early whitish gastric neoplasms can be easily misdiagnosed; differential diagnosis of gastric whitish lesions remains a challenge. We aim to build a deep learning (DL) model to diagnose whitish gastric neoplasms and explore the effect of adding domain knowledge in model construction. METHODS: We collected 4558 images from two institutions to train and test models. We first developed two sole DL models (1 and 2) using supervised and semi-supervised algorithms. Then we selected diagnosis-related features through literature research and developed feature-extraction models to determine features including boundary, surface, roundness, depression, and location. Then predictions of the five feature-extraction models and sole DL model were combined and inputted into seven machine-learning (ML) based fitting-diagnosis models. The optimal model was selected as ENDOANGEL-WD (whitish-diagnosis) and compared with endoscopists. RESULTS: Sole DL 2 had higher sensitivity (83.12% vs 68.67%, Bonferroni adjusted P = 0.024) than sole DL 1. Adding domain knowledge, the decision tree performed best among the seven ML models, achieving higher specificity than DL 1 (84.38% vs 72.27%, Bonferroni adjusted P < 0.05) and higher accuracy than DL 2 (80.47%, Bonferroni adjusted P < 0.001) and was selected as ENDOANGEL-WD. ENDOANGEL-WD showed better accuracy compared with 10 endoscopists (75.70%, P < 0.001). CONCLUSIONS: We developed a novel system ENDOANGEL-WD combining domain knowledge and traditional DL to detect gastric whitish neoplasms. Adding domain knowledge improved the performance of traditional DL, which provided a novel solution for establishing diagnostic models for other rare diseases potentially.

A deep-learning based system using multi-modal data for diagnosing gastric neoplasms in real-time (with video).

BACKGROUND: White light (WL) and weak-magnifying (WM) endoscopy are both important methods for diagnosing gastric neoplasms. This study constructed a deep-learning system named ENDOANGEL-MM (multi-modal) aimed at real-time diagnosing gastric neoplasms using WL and WM data. METHODS: WL and WM images of a same lesion were combined into image-pairs. A total of 4201 images, 7436 image-pairs, and 162 videos were used for model construction and validation. Models 1-5 including two single-modal models (WL, WM) and three multi-modal models (data fusion on task-level, feature-level, and input-level) were constructed. The models were tested on three levels including images, videos, and prospective patients. The best model was selected for constructing ENDOANGEL-MM. We compared the performance between the models and endoscopists and conducted a diagnostic study to explore the ENDOANGEL-MM's assistance ability. RESULTS: Model 4 (ENDOANGEL-MM) showed the best performance among five models. Model 2 performed better in single-modal models. The accuracy of ENDOANGEL-MM was higher than that of Model 2 in still images, real-time videos, and prospective patients. (86.54 vs 78.85%, P = 0.134; 90.00 vs 85.00%, P = 0.179; 93.55 vs 70.97%, P < 0.001). Model 2 and ENDOANGEL-MM outperformed endoscopists on WM data (85.00 vs 71.67%, P = 0.002) and multi-modal data (90.00 vs 76.17%, P = 0.002), significantly. With the assistance of ENDOANGEL-MM, the accuracy of non-experts improved significantly (85.75 vs 70.75%, P = 0.020), and performed no significant difference from experts (85.75 vs 89.00%, P = 0.159). CONCLUSIONS: The multi-modal model constructed by feature-level fusion showed the best performance. ENDOANGEL-MM identified gastric neoplasms with good accuracy and has a potential role in real-clinic.

Development and Validation of a Deep Learning-Based Histologic Diagnosis System for Diagnosing Colorectal Sessile Serrated Lesions.

OBJECTIVES: The histopathologic diagnosis of colorectal sessile serrated lesions (SSLs) and hyperplastic polyps (HPs) is of low consistency among pathologists. This study aimed to develop and validate a deep learning (DL)-based logical anthropomorphic pathology diagnostic system (LA-SSLD) for the differential diagnosis of colorectal SSL and HP. METHODS: The diagnosis framework of the LA-SSLD system was constructed according to the current guidelines and consisted of 4 DL models. Deep convolutional neural network (DCNN) 1 was the mucosal layer segmentation model, DCNN 2 was the muscularis mucosa segmentation model, DCNN 3 was the glandular lumen segmentation model, and DCNN 4 was the glandular lumen classification (aberrant or regular) model. A total of 175 HP and 127 SSL sections were collected from Renmin Hospital of Wuhan University during November 2016 to November 2022. The performance of the LA-SSLD system was compared to 11 pathologists with different qualifications through the human-machine contest. RESULTS: The Dice scores of DCNNs 1, 2, and 3 were 93.66%, 58.38%, and 74.04%, respectively. The accuracy of DCNN 4 was 92.72%. In the human-machine contest, the accuracy, sensitivity, and specificity of the LA-SSLD system were 85.71%, 86.36%, and 85.00%, respectively. In comparison with experts (pathologist D: accuracy 83.33%, sensitivity 90.91%, specificity 75.00%; pathologist E: accuracy 85.71%, sensitivity 90.91%, specificity 80.00%), LA-SSLD achieved expert-level accuracy and outperformed all the senior and junior pathologists. CONCLUSIONS: This study proposed a logical anthropomorphic diagnostic system for the differential diagnosis of colorectal SSL and HP. The diagnostic performance of the system is comparable to that of experts and has the potential to become a powerful diagnostic tool for SSL in the future. It is worth mentioning that a logical anthropomorphic system can achieve expert-level accuracy with fewer samples, providing potential ideas for the development of other artificial intelligence models.

Exploring the challenge of early gastric cancer diagnostic AI system face in multiple centers and its potential solutions.

BACKGROUND: Artificial intelligence (AI) performed variously among test sets with different diversity due to sample selection bias, which can be stumbling block for AI applications. We previously tested AI named ENDOANGEL, diagnosing early gastric cancer (EGC) on single-center videos in man-machine competition. We aimed to re-test ENDOANGEL on multi-center videos to explore challenges applying AI in multiple centers, then upgrade ENDOANGEL and explore solutions to the challenge. METHODS: ENDOANGEL was re-tested on multi-center videos retrospectively collected from 12 institutions and compared with performance in previously reported single-center videos. We then upgraded ENDOANGEL to ENDOANGEL-2022 with more training samples and novel algorithms and conducted competition between ENDOANGEL-2022 and endoscopists. ENDOANGEL-2022 was then tested on single-center videos and compared with performance in multi-center videos; the two AI systems were also compared with each other and endoscopists. RESULTS: Forty-six EGCs and 54 non-cancers were included in multi-center video cohort. On diagnosing EGCs, compared with single-center videos, ENDOANGEL showed stable sensitivity (97.83% vs. 100.00%) while sharply decreased specificity (61.11% vs. 82.54%); ENDOANGEL-2022 showed similar tendency while achieving significantly higher specificity (79.63%, p < 0.01) making fewer mistakes on typical lesions than ENDOANGEL. On detecting gastric neoplasms, both AI showed stable sensitivity while sharply decreased specificity. Nevertheless, both AI outperformed endoscopists in the two competitions. CONCLUSIONS: Great increase of false positives is a prominent challenge for applying EGC diagnostic AI in multiple centers due to high heterogeneity of negative cases. Optimizing AI by adding samples and using novel algorithms is promising to overcome this challenge.