Introduction

For training deep learning algorithms in medical imaging, application-specific data is often scarce. Deep learning systems are therefore generally pretrained on large publicly available labeled data sets of general imagery, unrelated to the envisioned application, to have the algorithm learn basic features from these widely available data followed by refinement training on the generally scarce application-specific images. Pretraining might be more effective if the images for pretraining resemble the envisioned application, i.e., domain-specific pretraining.

Aims & Methods

We investigated if pretraining on general endoscopic imagery results in a better performance of five existing AI systems with an application in gastro-intestinal endoscopy, compared to current state-of-the art pretraining approaches (i.e., supervised pretraining with ImageNet and semi-weakly supervised pretraining with the Billion-scale data set). Our group has created an endoscopy-specific dataset called GastroNet for pretraining deep learning systems in endoscopy. GastroNet consists of 5,084,494 endoscopic images retrospectively collected between 2012 and 2020 in seven Dutch hospitals. We created four pretrained models: one using GastroNet and three using ImageNet and/or the Billion-scale data set. The pretraining method was either supervised, self-supervised, or semi-weakly supervised. The pretrained models were subsequently trained towards five independent, commonly used applications in GI endoscopy, using their original application-specific datasets. Outcome parameters were: 1) classification and/or localization performance of the five trained applications; 2) change in performance when the number of available application-specific training data was reduced, to investigate a possible difference in performance drop for the different pretrained models.

Results

Overall, the domain-specific pretrained model resulted in a statistically superior performance for the five different GI applications. More detailed results are presented in Table 1. The superiority was also reflected in a smaller drop in performance when the number of application-specific training data were reduced artificially.

Table 1. Overview of performance of the five different application-specific data sets using four different pretrained models. Cells highlighted in green represent the highest scoring pretrained model per application-specific data set.

Conclusion

Domain-specific pretraining, using unlabeled general endoscopic images, is superior to current state-of-the-art pretraining approaches for developing deep learning algorithms in GI endoscopy. It also allows more effective use of the generally scarce application-specific endoscopy images. These findings might cause a paradigm shift in the development of AI systems in endoscopy.

Introduction

Initial reports have described rectal endoscopic knife-assisted full-thickness resection (kFTR) as a novel technique for achieving R0 resection in deeply invasive or technically challenging rectal lesions, avoiding the significant morbidity and mortality of rectal surgery.^1,2​​​​ We aimed to evaluate the feasibility, safety, and technical outcomes of rectal kFTR.

Aims & Methods

In this retrospective international study, we included consecutive patients who underwent rectal kFTR at 15 centres across 4 continents between April 2017 and April 2025. We included patients with adenocarcinoma suspected of deep submucosal invasion, severe fibrosis, or subepithelial lesions were included. Primary outcomes were technical success, en-bloc and R0 resection rates and adverse events. Secondary outcomes included rectal preservation, length of hospital stay, and recurrence. Risk factors for adverse events were evaluated using univariate and multivariate logistic regression.

Results

A total of 94 patients (mean age 65 years [IQR 58-73]; 56.4% male; 79.8% ASA I/II) underwent rectal kFTR. Median lesion size was 35 mm [IQR 25-50]. kFTR was performed for suspected deep invasion (57; 60.6%), severe fibrosis (31; 33%), and subepithelial lesions (6; 6.4%). Of lesions included, 59 (62.8%) were in the distal rectum, and 13 (13.8%) were anterior.

Technical success was achieved in 92 procedures (97.9%). Mean procedure time was 135 minutes (IQR 100-180). Antibiotics were administered in 83 procedures (88.3%) for a median of 5 days (IQR 5-5). Median hospital stay was 1 day (IQR 1 - 3). Overall, 19 adverse events occurred in 15/94 (16%) patients: 12 (63.1%) mild, 6 (31.6%) moderate, and 1 (5.3%) severe per ASGE lexicon. Delayed bleeding occurred in 4 (4.3%) patients, and post-procedural pain in 7 patients (7.4%). Surgery was required in one case for a recto-vaginal fistula. Increasing lesion size was the only risk factor independently associated with a higher risk of adverse events (OR 1.05, 95% CI 1.02–1.09, p = 0.003).

En-bloc and R0 resections were achieved in 91/94 (96.8%) and 74 (78.7%) procedures, respectively. Rectal preservation was achieved in 68 patients (72.3%) while, 25 (26.6%) underwent completion surgery for non-curative cancer resections, with no residual tumor in the surgical specimen in 11/25 cases (44%).

Among 59 patients with adenocarcinoma and available follow-up (median 12 months (IQR 6 - 23.5)), 3 (5.1%) experienced loco-regional recurrences, 1 (1.7%) had distant metastasis from a pT2 adenocarcinoma with surgery after kFTR, and 1 died of non-cancer-related causes.

Conclusion

kFTR appeared to be a feasible and safe technique for excising deeply invasive adenocarcinomas and complex fibrotic or subepithelial rectal lesions in expert centres. The high rate of R0 resections enabled accurate histopathological risk stratification, supporting individualised post-procedural management and rectal preservation in most patients. Rectal kFTR may serve as an alternative to device-assisted endoscopic full-thickness resection and local surgical excision, potentially avoiding radical surgery in highly selected patients. Further prospective research is needed to confirm these findings.

References

1. Argenziano ME*, Sorge A*, Poortmans PJ, Montori M, Balducci D, Hoorens A, Maroni L, Tate DJ, Knife-Assisted Full-Thickness Resection Guided by Pocket Detection Method for Detection and Complete Excision of Deeply Invasive Rectal Cancer, VideoGIE (2024), doi: https://doi.org/10.1016/j.vgie.2024.11.004.
2. Argenziano, M.E.*, Sorge, A.*, Hoorens, A., Montori, M., Poortmans, P.J., Smeets, S., Tornai, T., Debels, L.K., Desomer, L. and Tate, D.J. (2025), Knife-assisted full-thickness resection guided by the pocket-detection method for posterior deeply invasive rectal cancer: A novel endoscopic approach (with video). DEN Open, 5: e70116. https://doi.org/10.1002/deo2.70116

Introduction

Antireflux surgery (fundoplication) may counteract gastroesophageal reflux of all types of carcinogenic gastric contents, not only acid, and reduce esophageal acid exposure to a greater extent than antireflux medication (proton pump inhibitors). We therefore examined the hypothesis that antireflux surgery prevents esophageal adenocarcinoma to a larger degree than antireflux medication in high-risk patients of esophageal adenocarcinoma, i.e. those with Barrett’s esophagus.

Aims & Methods

This multi-national and population-based cohort study included all patients with a recorded diagnosis of Barrett’s esophagus from Denmark (in 2012-2020), Finland (1987-1996 and 2010-2020), Norway (2008-2020), or Sweden (2006-2020) in a national patient registry. Patients who underwent antireflux surgery were compared with non-operated patients using antireflux medication. The risk of esophageal adenocarcinoma was calculated using multivariable Cox regression, providing hazard ratios (HR) with 95% confidence intervals (CI) adjusted for age, sex, country, calendar year, and comorbidity.

Results

The cohort consisted of 33,939 patients with Barrett’s esophagus. Of these, 542 (1.6%) had undergone antireflux surgery. During up to 32 years of follow-up, the overall HR was not decreased in patients having undergone antireflux surgery compared to non-operated patients using antireflux medications, but rather increased (adjusted HR 1.9, 95% CI 1.1-3.5). The point estimates did not decrease with longer follow-up, but instead increased for each follow-up category, from 1.8 (95% CI 0.6-5.0) within 1-4 years of follow-up to 4.4 (95% CI 1.4-13.5) after >10 years of follow-up.

Conclusion

Antireflux surgery does not seem to be superior to antireflux medication in the prevention of esophageal adenocarcinoma among patients with Barrett’s esophagus.