UEG's content platform for gastroenterology education bringing recordings, courses, guidelines, and more together in one place.

Re-live UEG Week 2026 on Gutflix Gut Guide online What's new on Gutflix

Trending on Gutflix

Rate for better recommendations!

Your feedback helps us show you more of what you like and less of what you dislike!

More of this!

We will show you more similar content.

Not for me!

We won't suggest this again.

Bookmark!

Save this content for later!

New to myUEG? Create an account! Log In with myUEG account

UEG Poster

Standard Poster

This content is part of Gutflix. Log in with your myUEG account, or create one free, to watch it.

Not sure what you can access? Learn more about account types.

EVALUATION OF AN IMPROVED COMPUTER AIDED DETECTION SYSTEM FOR BARRETT’S NEOPLASIA ON REAL-WORLD IMAGING CONDITIONS

Rixta van Eijck van Heslinga 1, Martijn R. Jong 1, Carolus Kusters 2, Tim Boers 2, Joost van der Putten 2, Lucas Duits 1, Roos E. Pouw 1, Bas L.A.M. Weusten 3, Martin Houben 4, Alaa Alkhalaf 5, Fons van der Sommen 2, Peter de With 2, Jeroen de Groof 1, Jacques J. Bergman 1

Affiliations

1 Amsterdam UMC, Amsterdam, Netherlands

2 Technische Universiteit Eindhoven, Eindhoven, Netherlands

3 St Antonius Hospital, Nieuwegein, Netherlands

4 HagaTeaching Hospital, The Hague, Netherlands

5 Isala Dept. of Gastroenterology, Groningen, Netherlands

Summary

AI Generated

Summary is not available for this content yet.

Download PDF

Was this helpful?

Thanks for your feedback.

This summary was generated by an AI large language model based on the content transcript. It is for informational purposes only and should not be considered a substitute for clinical judgment. Always rely on your professional expertise and the full clinical context when making clinical decisions.

Abstract

Introduction

Computer Aided Detection (CADe) systems may improve detection of (pre-)malignant lesions by endoscopists. Most CADe systems are developed using homogeneous datasets collected by experts, which do not represent the heterogeneous imaging conditions of daily clinical practice. Such heterogeneity is caused by endoscopist-dependent variation (e.g. skill and experience) as well as endoscopist-independent variation (e.g. endoscopy equipment). Recently, we presented a CADe system for Barrett’s neoplasia¹. However, all training and test data originated from expert centers.

Aims & Methods

We aimed to develop an updated CADe system (CADe 2.0) demonstrating more robustness to the varied imaging conditions found in routine patient care.
Our previously published CADe system (CADe 1.0) was trained and internally validated on a high-quality dataset originating from 15 expert centers. The dataset included 6.337 neoplastic (1.362 patients) and 7.695 non-dysplastic images (1.139 patients).
The updated CADe system (CADe 2.0) underwent four improvements. First, we included more heterogeneous training data from our training set patient population. We extracted video frames from videos comprising various quality artefacts (e.g. blurred lenses or insufficient mucosal cleaning). Second, we applied advanced data augmentation methods to address the diversity of post-processing enhancement settings used by modern endoscopy systems. Third, we updated the CADe system with a Vision Transformer architecture², which has shown to improve performance on lower quality endoscopic images³. Finally, we transitioned our pretraining approach from semi-supervised to a self-supervised learning framework⁴.
Both CADe systems were evaluated using three prospectively collected, independent test sets. The first test set represented the true diversity and complexity encountered in daily clinical practice. This dataset comprised all videos captured from all included patients in five referral centers between April 2022 and January 2024 (431 ten-second BE videos from 115 patients). The second test set focused on endoscopist-dependent variation. It included three paired subsets with high, moderate and low-quality images from 122 patients. Quality was determined by factors such as esophageal expansion and mucosal cleaning. The third test set addressed endoscopist-independent variation. It contained 16 paired subsets of 396 images (122 patients). Every subset was based on a different post-processing enhancement setting of the endoscopy system.

Results

For test set 1, CADe 1.0 reached a sensitivity and specificity of 84% and 73%. CADe 2.0 scored 93% and 76%. For test set 2, CADe 2.0 reached higher sensitivity and specificity scores on high-quality images compared to CADe 1.0. In addition, CADe 2.0 performance loss on moderate and low-quality images was smaller. On test set 3 CADe 1.0 scores substantially varied depending on the test sets enhancement setting (82-94% sensitivity, 52-83% specificity). Variation of CADe 2.0 was limited to 91-95% and 81-88%.

Table 1. Results of CADe 1.0 and CADe 2.0

Test Set	Metric	CADe 1.0	CADe 2.0
Test set 1: Video Performance	Sensitivity Specificity	84% 73%	93% 76%
Test set 2: High Quality Images Test set 2: Moderate Quality Images Test set 2: Low Quality Images	Sensitivity Specificity Sensitivity (loss) Specificity (loss) Sensitivity (loss) Specificity (loss)	93% 68% 76% (-17%) 69% (+1%) 61% (-32%) 76% (+8%)	96% 85% 85% (-11%) 87% (+2%) 72% (-24%) 90% (+5%)
Test set 3: Enhancement Setting	Sensitivity range (min-max) Specificity range (min-max)	12% (82-94) 31% (52-83)	4% (91-95) 8% (81-88)

Conclusion

The updated CADe system displayed improved performance for detection of Barrett’s neoplasia in videos. More importantly, this CADe system also displayed less performance loss when exposed to lower image quality and less performance variability when confronted with differing post-processing enhancement settings.

References

1. Fockens N, Jong MR, Jukema JB, Pouw RE, Duits LC, Van Munster SN, et al. A deep learning system for detection of early Barrett's neoplasia: a model development and validation study. The Lancet Digit Health. 2023;5:905-21.
2. Yu W, Si C, Zhou P, Luo M, Zhou Y, Feng J, et al. MetaFormer Baselines for Vision2022 October 01, 2022:[arXiv:2210.13452 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2022arXiv221013452Y.
3. Kusters CHJ, Boers TGW, Jaspers TJM, Jukema JB, Jong MR, Fockens KN, et al., editors. CNNs vs. Transformers: Performance and Robustness in Endoscopic Image Analysis. Applications of Medical Artificial Intelligence; 2024 2024//; Cham: Springer Nature Switzerland.
4. Caron M, Touvron H, Misra I, Jégou H, Mairal J, Bojanowski P, et al. Emerging Properties in Self-Supervised Vision Transformers2021 April 01, 2021:[arXiv:2104.14294 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2021arXiv210414294C.

Conference

UEG Week Vienna 2024

Topics

Oesophagus

Submission format

Abstract

Session

Barrett's oesophagus (Posters)

Citation

United European Gastroenterology Journal 2024; 12 (Supplement 8)

Published

2024

More Like This:

Pancreatic resection: How much to cut?

Ihsan Ekin Demir

QUALITY OF LIFE AFTER SURGICAL RESECTION OF PANCREATIC CYSTIC NEOPLASMS: RESULTS OF THE PROSPECTIVE METAPAC STUDY

Bettina Oehrle

EVALUATION OF AN IMPROVED COMPUTER AIDED DETECTION SYSTEM FOR BARRETT’S NEOPLASIA ON REAL-WORLD IMAGING CONDITIONS

Fons van der Sommen

Mistakes in endoscopic retrograde cholangiopancreatography and how to avoid them

Mathieu Pioche

THE ENHANCED RECOVERY AFTER SURGERY PROTOCOL TAKE ITS PLACE IN INFLAMMATORY BOWEL DISEASE SURGERY

Giuliano Barugola

ERPC

Thierry Ponchon

	Patients under BTS n=67	Patients under ES n=69	T/χ2	p value
Intraoperative complication, no. (%)	2（3.0%）	9（13.0%）	4.626	0.031
ICU requirement, no. (%)	7（10.4%）	19（27.5%）	6.420	0.011
Temporary stoma, no. (%)	23(34.3%)	57(82.6%)	32.713	＜0.001
Permanent stoma, no. (%)	9(13.4%)	37/63(58.7%)	29.138	＜0.001
Postoperative hospitalization, day (mean±SD)	12.4±11.0	11.1±5.8	0.825	0.411
Total cost, ¥ (mean±SD)	63360.0±37302.9	56315.2±31224.4	1.180	0.240
Clinical success, no. (%)	49(73.1%)	44(63.8%)	1.379	0.240
30-day mortality, no. (%)	0	6（8.7%）	/	0.028
Overall mortality, no. (%)	29/54（53.7%）	38/59（64.4%）	1.338	0.247