Introduction

The European Union has signalled gastric cancer screening and prevention as a major health priority. The TOGAS consortium (TOwards GAstric cancer Screening in the European Union; togas.lu.lv), funded in response to this call, investigates different strategies for primary and secondary prevention of gastric cancer in Europe. One pilot study aims at determining the yield of having an oesophagogastroduodenoscopy (OGD) on the same day of a colonoscopy for colorectal cancer screening or polyp surveillance.

Aims & Methods

This is the first interim data on n=602 individuals recruited in 8 centres across Europe (France, Germany, Ireland, Latvia, Lithuania, the Netherlands, Portugal, and Spain). Individuals of the age of 50-74 years, attending for screening or surveillance colonoscopy, were invited to participate if they did not have a past diagnosis of gastric preneoplastic conditions requiring surveillance. Standardization of endoscopic procedures was assured by a rigorous protocol including all quality parameters defined by ESGE. This included standards for image documentation and biopsy sampling. Demographics, comorbidities, medication and the previous surgical and family history were recorded. Patients with marked preneoplastic changes (i.e. glandular atrophy or intestinal metaplasia [IM] at OLGA/OLGIM stage 3 or 4, and/or EGGIM score >5) were registered as requiring further endoscopic surveillance. Mann-Whitney U-test was applied for comparison of age between groups and Chi2 test for comparison of categorical variables (significance for p<0.05).

Results

The mean age of participants was 62 years (SD +/- 8.0), with 55.8% being male. Of 245 individuals who underwent an OGD prior to inclusion, 93.1% of these procedures were done >3years before and 60.4% were done at a tertiary centre. We identified 6 (1.0%) malignant/neoplastic lesions, 4 gastric cancers/dysplastic lesions and 2 oesophageal carcinomas. A positive H. pylori status was diagnosed in 34.2%. Endoscopically, a diagnosis of preneoplastic conditions was made in 217 participants (36.0%). Of these, 84 (14.0%) showed gastric atrophy, 45 (7.5%) gastric IM and 75 (12.5%) both conditions combined. Of those with gastric atrophy, this was classified as multifocal in 44 (27.6%) individuals, of those with IM in 45 (37.5%). OLGA stages 3 and 4 were confirmed for 1.9% and 0.8%, OLGIM stages 3 and 4 for 2.3% and 0.4%. Benign ulcers were diagnosed in 28 (4.7%) participants. The majority being gastric ulcers in 19 (3.2%) individuals, duodenal ulcers in 9 (1.5%), and oesophageal ulcers in 2 (0.3%). An additional 13 (2%) participants were diagnosed with Barrett’s oesophagus. It is of note that also 12 (2.0%) colorectal cancers were identified in the cohort. A positive H. pylori status was associated with the need for gastric surveillance (p=0.031).

Conclusion

This interim analysis of the first European pilot study on a combined screening for upper and lower GI tumours shows that this approach is feasible. Regarding the detection of upper GI malignancy and preneoplastic conditions, our findings are in line with previous systematic reviews, but we also demonstrate a reasonable rate of non-malignant findings that are also of relevance when defining the best screening strategy in Europe.

Disclosure

JB: Advisory fee from Flynn Pharma Ltd, Juvise Pharmaceuticals.
MCWS: Research Support from Sysmex, Norgine, Sentinel
OK: Speakers fee, advisory fee from Abbvie, Pfizer, Takeda, Janssen, Shire.

Introduction

Deep learning systems in gastrointestinal endoscopy are traditionally pretrained on large, diverse datasets comprising generic, non-medical images before being trained on scarce endoscopic data. Foundation models are models pretrained based on domain-specific datasets. For endoscopy, foundation models require large datasets of general endoscopic images. Yet datasets for developing such models remain limited. In this study, we present GastroNet-5M, a dataset comprising 4,820,653 endoscopic images of approximately 500,000 procedures.

Aims & Methods

GastroNet-5M is composed of anonymized endoscopic images collected from eight Dutch hospitals between 2012 and 2020. Using a self-supervised learning approach, this dataset enabled the development of a foundation model for various endoscopic AI applications. In this study, we compared GastroNet-5M-pretrained models against current state-of-the-art models with traditional pretraining approaches and all publicly available endoscopic foundation models.
We evaluated diagnostic performance on 9 classification and 9 segmentation tasks, including Barrett’s neoplasia detection, colorectal polyp characterization, and gastric cancer invasion depth prediction. All models were trained from scratch for each task. Data efficiency was assessed by repeating all classification tasks with progressively smaller training sets (i.e. 100%, 50%, 25%, 10% and 5%). Robustness was evaluated across 12 datasets reflecting data heterogeneity such as variation in endoscope manufacturer, imaging modality (e.g., virtual chromoendoscopy), and image quality.

Results

GastroNet-5M-pretrained models outperformed all benchmarks in 8 of 9 classification tasks, with an average AUC improvement of 6.8% (Table 1). For segmentation, they achieved higher Dice scores compared to all other models across all 9 tasks, with an average gain of 15.4%. In data efficiency experiments, GastroNet-5M models achieved the highest AUC in 5 of 9 tasks when training data was limited, while no other model led in more than one. In robustness testing, they outperformed other models in 8 of 12 tasks evaluating heterogeneous data conditions.

Table 1. Area-Under-the-Curve scores of the GastroNet-5M-pretrained model and benchmark models for all classification tasks.

Conclusion

GastroNet-5M, a multicenter dataset of over 5 million unlabeled endoscopic images, offers a valuable resource for pretraining deep learning models in endoscopy. The use of GastroNet-5M enhances diagnostic accuracy, reduces the need for scarce application-specific endoscopic imagery, and improves robustness of endoscopic AI systems to the inevitable data heterogeneity in clinical practice. GastroNet-5M is publicly accessible for further research and development.

References

1. Sanderson E, Matuszewski BJ. arXiv preprint arXiv:2401.06278. 2024.
2. van der Putten J, de Groof J, Struyvenberg M, et al. Artif Intell Med. 2020;107.
3. Wang Z, Liu C, Zhang S, et al. arXiv preprint arXiv:2306.16741. 2023.
4. Liu Z, Mao H, Wu C-Y, et al. arXiv preprint arXiv:2201.03545. 2022.
5. Dosovitskiy A, Beyer L, Kolesnikov A, et al. arXiv preprint arXiv:2010.11929. 2020.
6. Oquab M, Darcet T, Moutakanni T, et al. arXiv preprint arXiv:2304.07193. 2023.
7. Jocher G, Qiu J. Ultralytics YOLOv11, v11.0.0. 2024.
8. de Groof AJ, Struyvenberg MR, van der Putten J, et al. Gastroenterology. 2020;158(4):915–29.e4.
9. Fockens KN, Jukema JB, Boers T, et al. United European Gastroenterol J. 2023;11(4):324–36.
10. Jukema J, Kusters C, Jong M, et al. Endoscopy (ESGE Days). 2023.
11. GIANA Challenge. Available from: https://giana.grand-challenge.org
12. Borgli H, Thambawita V, Smedsrud PH, et al. Sci Data. 2020;7(1):283.
13. Polat G, Kani HT, Ergenc I, et al. Inflamm Bowel Dis. 2023;29(9):1431–9.
14. van der Zander QEW, Schreuder RM, Fonollà R, et al. Endoscopy. 2021.
15. Tang D, Zhou J, Wang L, et al. Front Oncol. 2021;11:654093.
16. Houwen BB, Hazewinkel Y, Giotis I, et al. Endoscopy. 2022;54(Suppl 1):eP133.

Introduction

Corpus atrophic gastritis (CAG) requires endoscopic-histological surveillance due to the risk of developing gastric neoplastic lesions (GNL). This study aimed to identify variables associated with GNL development at long-term follow-up using a feature ranking method based on a novel One-Class Support Vector Machine (SVM) model.

Aims & Methods

A dataset containing clinical, endoscopic, and histological variables from consecutive CAG patients (2001-2023) adhering to a surveillance program was considered. GNL presence at the longest available follow-up was recorded. Gastric biopsies and histological evaluations were conducted according to the updated Sydney system. CAG patients with any missing data or a GNL diagnosis at the baseline gastroscopy were excluded from the final analysis. A Fisher score-based feature ranking method and One-Class SVM were used to select the optimal subset of baseline variables associated with GNL development, aiming to achieve average sensitivity (Se) and specificity (Sp) above 80%.

Results

Overall, 355 CAG patients were initially considered. Of these, 36 were excluded due to the presence of GNL at baseline gastroscopy, and 216 for missing data. Thus, a total of 103 patients were considered and grouped into: CAG patients with [22 patients (F 68.1%), median age 68 (35-83) years] and without GNL at follow-up [81 patients (F 72.8%), median age 59(26-84) years]. After a median follow-up of 60 (12-192) months, 13 epithelial GNL (gastric adenocarcinoma or high/low-grade dysplasia) and nine type-1 gastric-neuroendocrine-tumors (T1gNET) were recorded. Parietal-cell-antibodies and pepsinogen-I < 30 μg/l were associated with both epithelial GNL and T1gNET. Antral inflammation and age > 60 were linked to epithelial GNL, while anti-thyroperoxidase-antibodies, smoking, and dyspeptic symptoms at CAG diagnosis were linked to T1gNET. Low-dose aspirin and H. pylori eradication therapy showed inverse associations with epithelial GNL and T1gNET, respectively.

Conclusion

This is the first study in which an AI model simultaneously considers clinical, endoscopic, and histological characteristics from a dataset of CAG patients, showing good sensitivity and specificity in identifying those variables associated with developing GNL. Nonetheless, we acknowledge the need for improved variable standardization and data balancing between groups to enhance model performance further.

Introduction

Histological assessment in Inflammatory Bowel Disease (IBD) is often required with endoscopic diagnoses to establish disease activity. This is evident in patients where histological healing is a major therapeutic goal and when endoscopic healing can also exhibit microscopic disease. IBD can manifest in numerous landmark features which are structural and inflammatory related. Histological activity is determined through erosion/ulceration, architectural changes in tissue morphology, mucin depletion and presence of varied inflammatory cells the colonic crypts, Muscularis Mucosae (MM) and Lamina Propria (LP). Histological scoring systems vary in complexity and methodology and subject to discrepancies in inter-rater consensus. Our model aims to provide accurate measurements of key histological features in Ulcerative Colitis (UC) used in scoring systems as both a decision support tool for improving pathologist scoring alignment and sensitive quantitative metrics of inflammatory cells to evaluate disease activity.

Aims & Methods

We developed an ensemble of convolutional neural networks to identify cell types, tissue foreground, crypts, MM, and haemorrhage in H&E images of IBD biopsy data. Models were trained with expert-labelled annotations comprising 34,289 image-annotation pairs (80% training, 20% validation) from in-house and publicly available data. Model accuracy was assessed with 6,464 held-out image-annotation pairs from 26 Whole Slide Images (WSIs) from all sites of the large intestine commonly sampled during endoscopic investigation, while ground truth annotations were reviewed by GI-trained pathologists to certify accuracy. We used geospatial methods to derive boundaries of the LP from other predicted domains, associate cells to delineated regions, measure architectural changes and derive spatially differentiated metrics of inflammatory burden in 66 patients with confirmed UC.

Results

The AI model spatially differentiates and segments cell types and tissue compartments. This produces metrics for inflammatory cell counts in regions of interest such as crypts, LP, and MM. Goblet cell segmentation within crypts allows for measurement of mucin depletion. This can be visualized for pathologists where each inflammatory cell type across the WSI is represented as density metrics. Automated quantification of inflammatory cells across AI-identified tissue domains throughout the bowels of UC patients revealed large intra- and inter-patient variability.

Conclusion

Our AI model can measure histological healing by detecting architectural change, inflammatory cell number, goblet cell reduction and erosions and ulceration. This allows for metrics and visual overlays to assist with pathologist reading and quantifying changes in mucosal biopsies in a clinical trial setting which requires extremely sensitive tools only AI can provide.

Disclosure

DW/AM/RL/PA/PW and CL are employees at Perspectum Ltd. EF and RG are consultants for Perspectum Ltd. All other co-authors have no conflicts of interest to declare relevant to this work.