Introduction

Guidelines reflect the latest, evidence based consensus for our clinical routine. The existing recommendations are numerous and subject to frequent change. Current learning methods are predominantly non-interactive. Although case-based learning is recognized as an cost effective educational tool [1, 2], it is often constrained by the limited number of available up-to-date case examples. To our knowledge, there is no reproducible and easily maintainable source for case-based training on guideline content.

Aims & Methods

We aimed to develop a versatile framework for the procedural generation of cases that represent a set of guideline recommendations in a structured format.
For this prototype, we evaluated recommendations for the pre-endoscopic assessment of risks for major bleeding or thromboembolic events based on the German guideline for quality in gastrointestinal endoscopy [3] (Chapters 3.2.2 to 3.2.5) to identify necessary information for database population (e.g., interventions with high or low bleeding risks). A list of case templates was then created, each consisting of a selected guideline excerpt along with a set of predefined properties inherited by the resulting cases. Undefined parameters were generated using random values reflecting unremarkable findings to introduce diversity.
The platform, including the database definition, was developed as a Python web application using the Django framework. Pre-defined database content such as entries for diseases, classifications, interventions, and guideline excerpts are stored in easily extendable or modifiable ".yaml" files.

Results

We established a structured database which stores the data required to determine risk of thrombembolic event. This includes pre-existing diseases (e.g., chronic kidney disease (CKD), atrial fibrillation) with corresponding classifications (e.g., CKD Stage 1 - 5), laboratory values (e.g., hemoglobin, eGFR), and relevant previous events (e.g., stroke, coronary in-stent stenosis). Furthermore, the database contains data to determine bleeding risk including different interventions (e.g., polypectomy, EMR, ESD, eFTRD, ultrasound guided biopsies as well as patient factors like liver cirrhosis) and lesions. Lesions like polyps may also be specified using morphology (Paris or lateral spreading type classification), location (organ and organ part), and size.
The resulting Python-based Django web application can repeatably generate lists of different cases representing the same set of predefined guideline recommendations. Moreover, future changes can be easily adapted by modifying the implemented rules or guideline texts.
Based on the desired focus - for example apixaban intake and different stages of CKD - the user may define a list of case templates which represent those aspects. Additionally, the user adds the desired questions for the trainee including corresponding guideline citations. Those cases may then be presented to trainees. In our prototype, the trainee was instructed to assess the risk of bleeding and thromboembolic events. The trainee was also asked to determine the pausing of each drug before the procedure if it is necessary. After completing a case, we present the relevant guideline excerpts, including exact citations, to the trainee.

Conclusion

Our platform can be adapted to other domains and future guideline updates can be incorporated with minimal effort, making this tool suitable for diverse case-based learning or for evaluating knowledge in a consistent, benchmark-like manner.

References

[1] Thistlethwaite, J. E., et al. (2012). The effectiveness of case-based learning in health professional education. A BEME systematic review: BEME Guide No. 23. _Medical Teacher_, _34_(6), e421–e444. https://doi.org/10.3109/0142159X.2012.680939
[2] McLean, S. F. (2016). Case-Based Learning and its Application in Medical and Health-Care Fields: A Review of Worldwide Literature. Journal of Medical Education and Curricular Development, 3. https://doi.org/10.4137/JMECD.S20377
[3] Denzer, E. et al. (2015) S2k guideline: quality requirements for gastrointestinal endoscopy, AWMF registry no. 021–022, http://dx.doi.org/10.1055/s-0041-109520

Introduction

Automated identification and discrimination of individual colorectal polyps through artificial intelligence (AI) is a crucial step towards the development of colonoscopy report generation systems. The basic information of a colonoscopy report should contain the total number of identified polyps and characteristic images of the polyps, which will determine the follow-up time and treatment. However, there is no widely adopted solution to accomplish this task.

Aims & Methods

In this study, we propose an AI-based pipeline for automatic individual polyp identification and retrieval of corresponding polyp images to simplify and accelerate colonoscopy report generation. A pipeline consisting of a Siamese Twin Neural Network and a hierarchical density-based spatial clustering method was implemented for individual polyp identification and discrimination. The AI model was trained and validated using frames containing polyps which were extracted from 549 and 61 full-length colonoscopies videos respectively, recorded in 7 different endoscopy centers. The pipeline was tested on frames containing a total of 29 different polyps extracted from 10 colonoscopies. The evaluation metrics were two measures to estimate the quality of clustering ranging from 0 (no correlation) to 1 (perfect correlation): normalized mutual information (NMI) and adjusted rand index (ARI). Additionally, we calculated the precision of corresponding individual polyp images.

Results

Testing on frames with visible polyps obtained from 10 colonoscopy examinations demonstrated that the pipeline achieved an NMI of 0.862 and an ARI of 0.835, indicating that at least 80% of the data can be accurately clustered and identified. The average precision was 0.763, representing the probability of retrieving the correct and corresponding individual polyp images.

Conclusion

We have developed a pipeline that uses artificial intelligence techniques to identify individual polyps and retrieve representative polyp images. These approaches have the potential to simplify and accelerate the process of automatically generating a colonoscopy report, while also providing characteristic images for each detected polyp for future use.

Introduction

In colonoscopy reports, the total number of polyps discovered and the characteristics of the polyp images are essential data in determining follow-up timing and treatment. Using artificial intelligence (AI) to differentiate individual colorectal polyps may potentially provide the needed information and reduce workload during examinations, especially in cases with multiple polyps. However, there is currently no widely adopted solution to accomplish this goal.

Aims & Methods

In this study, we built an AI-based pipeline to discriminate frames with polyps from colonoscopy videos. The pipeline consists of a Neural Network to compute features of each image and a hierarchical density-based spatial clustering method to group the features. Through this pipeline, we intend to count the number of polyps appearing in colonoscopy videos. The AI model was trained on 398 full-length colonoscopy videos with frames containing polyps and was validated on 44 videos. These videos were recorded in 7 different endoscopy centers from November 2021 to November 2022. The pipeline was tested in 50 colonoscopies, which contained a total of 83 different polyps. Two evaluation metrics are used: the error in the number of polyps, to quantify the disparity between the actual and predicted numbers of polyps; and precision, to evaluate the accuracy of retrieving individual polyp images.

Results

Testing on frames with visible polyps obtained from 50 colonoscopy examinations demonstrated that the pipeline achieved an average error in the number of polyps of 1.3. And the average precision was 0.66, representing the probability of retrieving the corresponding individual polyp images. Among these examinations, there are 14 examinations with more than one polyp, which have an average error in the number of polyps of 1.1 and an average precision of 0.55.

Conclusion

We have developed a pipeline that utilizes artificial intelligence techniques to identify individual polyps and retrieve representative polyp images. Despite the pipeline encountering challenges in cases with only one polyp, it still holds potential. Overall, this approach provides a promising direction for retrieving individual polyp images.

Disclosure

No

Introduction

Training with silicone models in colonoscopy is limited by the lack of realistic texture and appearance of mucosal tissue. Artificial intelligence (AI) offers a novel solution by enabling simulation of realistic appearing mucosa, which can enhance the overall visual realism of silicone-based models for endoscopic training.

Aims & Methods

We developed, using 239,875 images from 3,452 colonoscopies recorded using five different endoscopic processors, a style transfer AI model capable of transforming endoscopic video footage from a silicone colon into a video that visually mimics real mucosal tissue. The model was applied to two representative clips recorded from a silicone colon model, covering lesions and normal mucosa. Both the original and AI-enhanced videos were shown to N = 12 gastroenterologists (board certified and in training), who provided quantitative evaluations using a 5-point Likert scale (1: Strongly disagree, 5: Strongly agree) and free-text qualitative feedback. We conducted the statistical analysis using Wilcoxon signed-rank tests with Pratt adjustment for zero differences and analyzed the qualitative feedback with an inductive content analysis after Mayring.

Results

All physicians prefer the AI-enhanced videos over the original silicone footage due to the more realistic representation of the colon. This can also be seen from the quantitative data, as the participants rated the AI-enhanced videos as significantly more visually realistic (M = 2.42 vs 4.17, Z = -3.22, p < .001), helpful for training (M = 3.67 vs 4.25, Z = -2.65, p = .016) and polyp characterization (M = 2.92 vs 3.75, Z = -2.38, p = .027), favored for usage (M = 2.92 vs 4.00, Z = -2.80, p = .005), and facilitating optical diagnoses (M = 3.58 vs 4.42, Z = -2.23, p = .039). Though, the participants rated the original silicone footage as significantly less mentally demanding than the AI-enhanced videos (M = 1.67 vs 2.42, Z = -2.61, p = .016). Open-ended feedback highlighted the more realistic AI generated mucosa as an enhancement (100 %) to the unrealistic silicone surface (92 %) while suggesting that further development should focus on integrating more realistic lesions (33 %) and increasing mucosal details (42 %).

Conclusion

AI-driven simulation of normal mucosa improves the realism, acceptance, and educational value of silicone colon models used in endoscopy training.

Introduction

Measurement of the size of colorectal polyps during colonoscopy is carried out mainly by visual estimation. We developed a novel polyp size measurement system (Poseidon) based on artificial intelligence that uses the auxiliary water jet as a measurement reference.

Aims & Methods

Visual estimation, biopsy forceps measurement, and Poseidon were compared using a CT colonography-based silicon colon model with 28 polyps of known size. Four experienced endoscopists estimated polyp size visually and with biopsy forceps. The endoscopists took images of each polyp with the water jet in its proximity. These images were then analyzed by Poseidon to measure the size of the polyp.
Additionally, the Poseidon measurements were compared to visual size estimates of 29 colorectal polyps examined during clinical routine.

Results

In the silicone colon model, the visual estimation had the largest median percent error (PE) of 25.2% (interquartile range (IQR: Q1, Q3): 11.8, 43.1), followed by the biopsy forceps measurements with a median PE of 20% (IQR: 9.1, 34.3). Poseidon had a significantly lower median PE of 7.4% (IQR: 3, 15.7, p<0.001) than the other two methods.
At routine colonoscopies, Poseidon had a significantly lower median PE (7.7%; IQR: 2.7, 12.6) than the visual estimation (22.1%; IQR: 12.4, 34.1, p<0.001).

Conclusion

In this work, we present a novel method to measure the size of colorectal polyps, which has a significantly higher accuracy than other common polyp sizing methods.