Introduction

Accurate labeling of obstruction site on upright abdominal radiograph is a challenging task. The lack of ground truth leads to poor performance on supervised learning models. To address this issue, self-supervised learning (SSL) is proposed to classify normal, small bowel obstruction (SBO), and large bowel obstruction (LBO) radiographs using a few confirmed samples.

Aims & Methods

A total of 1500 confirmed and 1922 unlabeled radiographs were categorized based on the ground truth. The SSL model was firstly trained on the unlabeled radiographs, and then fine-tuned on the confirmed radiographs. ResNet50 and VGG16 were used for the embedded base encoders, whose weights and parameters were adjusted during training process. Furthermore, it was tested on an independent dataset, compared with supervised learning models and human interpreters. Finally, the t-SNE and Grad-CAM were used to visualize the model's interpretation.

Results

The proposed VGG16-backboned SSL model showed superior classification performance (accuracy 0.847, F1-score 0.846, MCC 0.873, Kappa 0.770), followed by the ResNet50-backboned model and senior radiologist. When recognizing SBO and LBO radiographs, the SSL model reached the best accuracy (0.907, 0.907) and F1-score (0.865, 0.860). Additionally, the SSL model exhibited better cluster performance through t-SNE and demonstrated outstanding lesion recognition ability through Grad-CAM.

Conclusion

The proposed SSL model exhibited outstanding performance in classifying bowel obstruction on upright abdominal radiograph, using unlabeled images and limited ground truth. Our study supports the feasibility of SSL model for computer-aided diagnosis of diseases that are difficult and time-consuming for labeling.

Introduction

Colorectal neuroendocrine tumors (NETs) and polyps have similar endoscopic manifestations and are easily misidentified. Due to the low incidence of NETs, it is difficult to obtain adequate images to fit deep-learning models. This study introduced a few-shot learning (FSL) model for the ternary classification of NETs and polyps on endoscopic images.

Aims & Methods

Three classes of endoscopic images (56 serrated lesions and polyps [SLPs], 86 traditional adenomas and 53 NETs) were collected from three datasets. First, three images from each class were randomly selected for Support Set (SS), while others for Query Set (QS). Second, ResNet50 V2 was pretrained based on ImageNet and then on esophageal endoscopic images Third, imaging eigenvectors were extracted based on a dual-pretrained model. Fourth, the Euclidean distances to the nine SS images were calculated for each QS image. Finally, a k-nearest neighbor classifier was used for classification based on Euclidean distances. The proposed FSL model was tested in three rounds.

Results

The performance of the FSL model (macro-area under the curve [AUC], 0.731; macro F1-score, 0.674; Matthews Correlation Coefficient [MCC], 0.526; Kappa, 0.523) was better than that of the traditional ResNet50 V2, junior endoscopist, and senior endoscopist. When identifying NETs, the proposed FSL model achieved the highest accuracy 0.823, sensitivity 0.653, precision 0.673 and F1-score 0.659, whereas the senior endoscopist achieved the highest specificity 0.904.

Conclusion

This study presented a FSL method based on transfer learning and metric learning to distinguish between SLPs, adenoma, and NETs on endoscopic images. This suggests that FSL might be an effective method for computer-aided diagnosis of rare diseases.

Introduction

Diabetic patients have a higher risk of colorectal cancer compared to non-diabetic patients, and studies suggest that this increased risk is associated with inadequate bowel preparation before colonoscopy.

Aims & Methods

This study aimed to develop machine-learning (ML) predictive models to based on diabetic patients for inadequate bowel preparation.
In this retrospective machine learning (ML) study, patients who met the diagnostic criteria for “Diabetes,” “Type 1 Diabetes,” “Type 2 Diabetes,” or “Diabetes with **” according to the International Classification of Diseases (ICD)-10 during a single hospitalization were included from the Wuxi Medical Center – Colorectal Adenoma Database (WMC-CAD). The database records patient admissions to the Wuxi Medical Center, affiliated with Nanjing Medical University, China, from January 1, 2018 to December 31, 2024 (including demographic characteristics, laboratory test results, and treatment regimens). Six ML algorithms were employed to predict inadequate bowel preparation in diabetic patients: logistic regression, support vector machine (SVM), gradient boosting machine (GBM), neural networks, XGBoost, and LightGBM. The area under the receiver operating characteristic curve (AUROC) was the primary metric for model performance assessment. SHapley Additive exPlanations (SHAP) values were calculated to evaluate the contribution of each feature in the highest-performing model.

Results

A total of 4,827 patients were included in the model development, with 3,863 patients in the training set and 964 in the validation set. In the training set, 815 patients (21.1%) had inadequate bowel preparation, while 3,048 (78.9%) had adequate preparation. Features with over 30% missing data were excluded, and missing values in the dataset were imputed using the k-nearest neighbors (KNN) method. Fifty-one variables were initially considered as candidate predictors.
The Boruta algorithm was employed for feature selection, identifying 19 variables. Subsequent Spearman correlation analysis eliminated strongly correlated variables to prevent model overfitting or instability. The remaining 17 variables underwent variance inflation factor (VIF) testing, confirming no significant multicollinearity, and were retained for model development.
The LightGBM model demonstrated the best predictive performance, achieving an AUROC of 0.779 (95% CI: 0.760–0.797) in the training set and 0.768 (95% CI: 0.749–0.786) in the internal validation set. LightGBM also attained the highest accuracy (65.6%) and the largest area under the precision-recall curve (AUPRC, 0.905) in the validation set. Its sensitivity and specificity were well-balanced at 63.0% and 75.2%, respectively.
SHAP analysis revealed that constipation and use of GLP-1 receptor agonists were the top two contributing factors to inadequate bowel preparation in diabetic patients.

Conclusion

LightGBM exhibited the best performance for predicting. SHAP values highlighted the importance of constipation and use of GLP-1 receptor agonists, for whom more polyethylene glycol or auxiliary medication should be used.

Introduction

Adequate bowel preparation (BP) is crucial for high-quality colonoscopy surveillance in ulcerative colitis (UC), because an inadequate one leads to incomplete colonoscopy and higher rate of missing precancerous lesions, which are usually flat or subtle in this setting. There is conflicting evidence regarding whether disease activity or extension are risk factors for inadequate BP in patients with inflammatory bowel disease (IBD); furthermore, data on the optimal PEG-regimen which should be used in this setting are limited. The aim of this study was to assess which factors influence the quality of BP in this specific setting.

Aims & Methods

This is a multicentre, cross-sectional, retrospective study carry out in four referral centres between January-2021 and December-2023, including quiescent UC patients with endoscopic remission (Mayo endoscopic score=0) who underwent colonoscopy. Bowel preparation quality was assessed using the Boston Bowel Preparation Scale (BBPS) and performed with 1L-PEG-ASC (Plenvu®) and 2L-PEG (Moviprep® and Clensia®), without any specific allocation or instructions. Demographic, anamnestic and endoscopic data were evaluated and analysed.

Results

Overall, 261 UC patients (F 99, 37.9%, mean age 55.6±14.5) were recruited. Patients with previous maximum extension of proctitis (E1) were 31 (12.1%), with left-sided colitis (E2) were 104 (40.6%) and with pancolitis (E3) were 121 (47.3%); additionally, patients with a history of advanced therapy were 104 (39.8%). Entire cohort adequate preparation rate was 90.4%.
1L-PEG-ASC preparation showed significantly higher adequate BP rate compared to 2L-PEG preparations (93.7% vs 85.4%, p=0.032). Right colon adequate BP did not differ among the BPs (93.7% vs 91.3%). Patient demographics, smoking status, maximum disease extension, history of advanced therapy or presence of pseudopolyps did not significantly affect BP quality (p>0.05).

Conclusion

In a relatively unbiased setting, such as quiescent UC patients in endoscopic remission, bowel preparation volume, specifically 1L-PEG-ASC, was the primary determinant of BP quality, with neither disease extent nor a history of advanced therapy affecting BP. These findings suggest that 1L-PEG-ASC preparation may be the preferred option for colonoscopy BP in UC patients when not contraindicated.

Introduction

Adhesive small bowel obstruction (ASBO) constitutes one of the most common causes of emergency room admissions worldwide. Although open adhesiolysis (OA) has long been the most widely adopted surgical approach, laparoscopic adhesiolysis (LA) is gaining momentum. However, current evidence remains controversial, as most of the studies suffer from a high risk of bias. This study aims to elucidate the comparative clinical outcomes and safety profiles of LA and OA in ASBO.

Aims & Methods

A systematic search of Medline, Embase, CINAHL, Cochrane Central Register of Controlled Trials, and Scielo databases was performed to identify eligible studies that compared LA and OA for ASBO published up until March 2023. Only randomized controlled trials (RCTs) and matched observational studies were included. The primary outcomes encompassed duration of surgery, postoperative complications, mortality, and postoperative hospital stay. We used random-effects models for data pooling. Heterogeneity amidst studies was evaluated via the I2 statistic.

Results

From 1955 articles screened, six studies evaluating a total of 1053 patients (mean age: 65 years; 67% females) fulfilled the selection criteria and were included. From these patients, 39% had an ASA classification ≥3 (95% CI 23%-57%. I2: 96%), 21% had two or more previous abdominal surgeries (95% CI 6%-42%. I2: 94%), 8% had evidence of irreversible bowel ischemia detected at surgery (95% CI 5%-12%. I2: 31%) and 48% (n=507) underwent laparoscopic lysis of adhesions. No significant differences regarding surgery duration (MD -5.52; 95% CI -15.98, 4.94. I2: 50%), iatrogenic bowel injury (OR 1.20; 95% CI 0.56, 2.55. I2: 0%), mortality (OR 1.30; 95% CI 0.46, 3.68. I2:0%), and the development of any postoperative complication (OR 2.61; 95% CI 0.19, 36.34. I2: 91.7%) were observed. However, LA was associated with a reduced postoperative stay compared to the open approach (MD -2.40; 95% CI -3.26, -1.54. I2: 12%).
Figure. Forest plot of postoperative hospital stay (days)

Conclusion

Although OA remains the standard procedure for the treatment of ASBO, LA appears to be a feasible approach to treat this condition in selected populations, with comparable results in terms of postoperative complications and mortality, while being associated with a significantly reduced postoperative length of stay. Additional RCTs with robust methodology should be performed in order to confirm these findings.

Disclosure

The authors have nothing to disclose.

Introduction

Inadequate bowel preparation is linked to reduced adenoma detection rates (ADR), and establishing a threshold for bowel preparation below which colonoscopies should be repeated early is crucial for patient outcomes. However, manual bowel cleansing efficacy evaluation relies on the clinical observer's experience, resulting in subjectivity with clinical and economic implications. The Boston Bowel Preparation Scale (BBPS) is the most validated scoring scale. Our group previously constructed an AI-based bowel preparation assessment model to calculate e-BBPS scores.

Aims & Methods

This prospective, single-center study was conducted at the Eighth Affiliated Hospital, Sun Yat-sen University in China from October 2021 to November 2022. Eligible patients underwent screening colonoscopy by one of three study endoscopists. The AI calculated the e-BBPS score during colonoscope withdrawal. The endoscopist assessed bowel preparation adequacy using BBPS after the colonoscopy. If both BBPS and e-BBPS deemed bowel preparation adequate, the same endoscopist performed a second colonoscopy. If either considered it inadequate, the patient underwent bowel re-cleansing before the second colonoscopy. Video recordings were evaluated for bowel preparation using BBPS by five independent endoscopist raters.

Results

From October 8, 2021, to November 9, 2022, a total of 393 patients were included for analysis. In patients with unqualified-AI, the >5mm AMR was significantly higher than in patients with qualified-AI (p=0.0056, OR 0.2734, 95% CI 0.1139, 0.6565), as were the AMR <0.001 (p=0.2532, 95% CI 0.1583, 0.4052), diminutive AMR (p<0.001, OR 0.2160, 95% CI 0.1206, 0.3867), >5mm PMR (p=0.0152, OR 0.4335, 95% CI 0.2288, 0.8213), and PMR (p <0.001, OR 0.4532, 95% CI 0.3447, 0.5958). No significant differences were observed in the overall >5mm ADR, ADR, >5mm PDR, and PDR. In the first pass, no statistical differences were found in ADR between the qualified-AI and unqualified-AI groups, though there was a trend toward a higher ADR in the qualified-AI group. In the second pass, patients in the qualified-AI group achieved significantly lower >5mm ADR, ADR, diminutive ADR, advanced ADR, >5mm PDR, and PDR compared to those in the unqualified-AI group.
Cohen's kappa analysis revealed high agreement between expert raters, clinical scoring, and e-BBPS (ranging from 0.7791 to 0.8322). However, the agreement between general endoscopists and other raters was only low to moderate (0.2937 to 0.6071). Fleiss' kappa also indicated that the overall agreement among all raters was only 0.575 (p < 0.001, 95% CI 0.574, 0.575), while the Fleiss' kappa for Expert and Clinical scoring, as well as Expert, Clinical scoring, and e-BBPS were 0.809 (p < 0.001, 95% CI 0.807, 0.810) and 0.800 (p < 0.001, 95% CI 0.799, 0.801), respectively.

Conclusion

This study validates the e-BBPS bowel preparation evaluation threshold, confirming that patients with a score of ≤3 have adequate bowel preparation for detecting >5mm adenomatous polyps, while those with a score of >3 require repeat bowel preparation. This conclusion supports the implementation and promotion of fully automated bowel preparation assessment systems, which may improve the objectivity of bowel preparation evaluation and standardize clinical practice for screening colonoscopies in the future.