Introduction

Early detection of neoplasia in Barrett’s Esophagus (BE) remains challenging. Although computer-aided detection (CADe) systems are increasingly available to support endoscopists, successful CADe implementation largely depends on the quality of endoscopic images, which can vary significantly in clinical practice. This study introduces an innovative computer-aided quality (CAQ) system designed for BE endoscopy, capable of assessing various image quality parameters, providing real-time feedback to the endoscopist, and operating in conjunction with an established CADe system.

Aims & Methods

The CAQ system was trained on 4,926 images from 243 BE patients to assess both objective (e.g., blurriness, illumination, and specular reflections) and subjective (i.e., mucosal cleanliness and esophageal expansion) image quality factors. The system is able to provide real-time feedback on these quality parameters to the endoscopist. Additionally, it is configured to function as a preprocessing filter for a previously developed CADe system, automatically excluding lower-quality images. To validate the CAQ system’s performance as a standalone tool, we constructed an endoscopic image quality test set, which included 647 images from 51 non-dysplastic BE patients across eight hospitals. These images were annotated for objective (e.g. motion blur, lighting) and subjective (i.e. esophageal expansion, mucosal cleanliness) image quality parameters by two experts. Additionally, to evaluate the CAQ system’s role as a preprocessing step for an established CADe system, we created the Barrett CADe test set, containing 956 frames from 62 neoplastic BE patients and 557 frames from 35 non-dysplastic BE patients from 12 Barrett referral centers. Outcome measures included agreement scores with experts for standalone image quality assessment and sensitivity, specificity, and Area-under-the-Curve (AUC) for BE neoplasia detection.

Results

As a stand-alone system, the CAQ system achieved Cohen’s Kappa scores for agreement with expert annotators of 0.73, 0.91, and 0.89 for objective image quality parameters, mucosal cleanliness, and esophageal expansion, respectively—comparable to inter-annotator scores of 0.73, 0.93, and 0.83. When used as a preprocessing filter for the CADe system, the CAQ system improved the CADe's sensitivity from 82% to 90% and the AUC from 87% to 91%, while specificity remained stable at 75% (Table 1).

Table 1. Performance of the CADe system for Barrett’s neoplasia detection with and without the CAQ system.

Conclusion

This study introduces the first CAQ system specifically aimed at automated quality control in BE endoscopy. As a standalone tool, the system effectively distinguishes poorly from well-visualized mucosa. When combined with an established CADe system, it enhances detection rates for neoplasia, highlighting its potential to improve diagnostic accuracy in BE surveillance.

Introduction

Timely detection of neoplasia in patients with Barrett’s esophagus (BE) has significant influence on patient outcome. However, endoscopic detection remains challenging due to its often subtle appearance. Computer Aided Detection (CADe) systems may assist in neoplasia detection.

Aims & Methods

We aimed to develop and externally validate a CADe system for real-time BE neoplasia detection on multiple video test sets and provide a benchmark performance by general and BE expert endoscopists.
The system underwent pretraining with ImageNet followed by domain-specific pretraining with GastroNet. GastroNet comprises >5 million endoscopic images from the gastrointestinal tract. The system was then trained and validated on a large, heterogeneous BE dataset originating from 15 international endoscopy centers, comprising 6.337 neoplastic (1.362) patients) and 7.695 non-dysplastic images (1.139 patients). All images had histopathological confirmation. Neoplastic images were delineated by expert endoscopists. The system was internally validated on a new, unseen dataset of 58 neoplasia patients (100 images and 68 videos) and 36 non-dysplastic patients (100 images and 58 videos).
For external validation, the system was tested on two independent, prospectively collected video test sets. The "all-comers test set" was representative for cases encountered in daily practice and contained all cases acquired by the 15 participating centers, from January 2022 to March 2022, comprising 71 neoplastic (45 patients) and 180 non-dysplastic (66 patients) videos. The "benchmarking test set" was artificially enriched with subtle neoplasia, representing the challenging cases in clinical practice. It comprised of 47 neoplastic (47 patients) and 141 non-dysplastic (82 patients) videos. This second test set was evaluated by 63 general endoscopists in 2 phases to provide a benchmark performance. In phase 1 endoscopists evaluated all cases without CADe assistance. After a six week wash-out period, endoscopists evaluated the same test set with CADe assistance. Finally, 14 internationally renowned BE experts evaluated this test set.

Results

Sensitivity and specificity of the CADe system were 97% and 85% for the all-comers test set and 91% and 82% for the benchmarking test set. Sensitivity of general endoscopists increased from 67% to 79% with CADe assistance, whilst specificity decreased from 96% to 94%. General endoscopists without CADe assistance were outperformed by CADe (p<0.001) and general endoscopists with CADe (p<0.001). The median sensitivity and specificity of experts were 86% and 90%. Non-inferiority of CADe to experts for neoplasia detection could be concluded.

Conclusion

CADe outperformed general endoscopist in terms of detecting subtle early BE neoplasia. Providing CADe assistance to general endoscopists significantly improves their detection of subtle neoplasia. CADe detects virtually all BE neoplasia encountered in daily practice. CADe has a detection performance which is non-inferior to international experts.

Disclosure

This study received financial and logistical support from Olympus Tokyo.

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.