A108 AUTOMATED DETECTION OF ILEOCECAL VALVE, APPENDICEAL ORIFICE, AND POLYP DURING COLONOSCOPY USING A DEEP LEARNING MODEL

Mahsa Taghiakbari, S Hamidi Ghalehjegh, E Jehanno, T Berthier, L di Jorio, Alan Barkun, Érik Deslandres, Sioui Maldonado Bouchard, Sacha Sidani, Yoshua Bengio, Daniel von RentelnJournal of the Canadian Association of Gastroenterology 2023 · DOI

TL;DR

This paper introduces a deep learning model capable of automatically identifying key anatomical landmarks (ileocecal valve and appendiceal orifice) and differentiating them from polyps and normal tissue during colonoscopies. This has the potential to significantly improve the efficiency and accuracy of colonoscopy reporting and quality assurance.

What Problem Does This Paper Address?

Ensuring the completeness of colonoscopy examinations is crucial, and this is often confirmed by documenting the identification of the ileocecal valve (ICV) and appendiceal orifice (AO). Currently, this process relies on manual photo-documentation and reporting by endoscopists, which can be time-consuming and prone to human error. The challenge is to automate this process, making it more reliable and efficient. This paper addresses the need for an AI-powered solution that can automatically identify these critical landmarks and distinguish them from other visual elements like polyps and normal colon mucosa, thereby enhancing the quality and consistency of colonoscopy reports.

Methods

The researchers developed a Deep Convolutional Neural Network (DCNN) model. They employed a transfer learning technique, which involves taking a pre-trained model and fine-tuning it on their specific dataset of endoscopic images. This approach leverages existing knowledge from models trained on large image datasets, accelerating the training process and often leading to better performance, especially when the target dataset is not massive. The model was trained using a cross-entropy loss function and an back-propagation algorithm, standard techniques for training neural networks for classification tasks. The model was designed to classify images into four categories: AO, ICV, normal mucosa, and polyps.

Datasets

The study utilized a dataset of full-length colonoscopy videos from 318 patients. From these videos, they extracted 25,444 unaltered frames, which were then divided into non-overlapping training, validation, and test sets. A significant effort was made to ensure data quality, with frames reviewed and annotated by a team of three clinicians. In total, 86,754 frames were examined, with specific counts for each class: 7982 for AO, 8374 for ICV, 32,971 for polyps, and 37,427 for normal mucosa. Importantly, only frames from white-light colonoscopies were used, excluding those from narrow-band imaging. The training set comprised 21,503 frames from 272 patients, while the validation and test sets contained 1,924 frames (from 25 patients) and 2,017 frames (from 21 patients), respectively. This represents a newly curated dataset for this specific task.

Key Insights and Ideas

  • AI as a "Second Pair of Eyes": The core idea is to use AI not to replace the endoscopist, but to augment their capabilities, ensuring critical landmarks are consistently identified and documented.
  • Leveraging Transfer Learning: The successful application of transfer learning highlights its effectiveness in adapting general image recognition models to the specific, often challenging, visual domain of endoscopy.
  • The Importance of High-Quality Annotation: The meticulous annotation process by multiple clinicians underscores the foundational role of accurate ground truth data in training robust deep learning models for medical applications.
  • Beyond Simple Detection: The model doesn't just detect landmarks; it differentiates them from other visual elements, which is crucial for accurate reporting and avoiding false positives.

Key Results

The DCNN model demonstrated strong performance:

  • It could identify both AO and ICV in 85.71% of patients (18 out of 21) when accuracies were above a 40% threshold, indicating reliable landmark identification without excessive false alerts.
  • The accuracy for differentiating AO from normal mucosa was 86.37%.
  • The accuracy for differentiating ICV from normal mucosa was 86.44%.
  • The model achieved an accuracy of 88.57% for differentiating polyps from normal mucosa.

These results suggest the model can reliably distinguish between the target anatomical landmarks and other common findings in colonoscopies.

Limitations and Future Work

The paper acknowledges that the primary outcome was defined as the proportion of patients where the AI model could identify both AO and ICV with accuracies above a 40% threshold. While this threshold is presented as a point where reliable identification can be assumed without increasing false-positive alerts, it's a relatively low bar for "reliable identification." Future work could focus on achieving higher accuracy thresholds and exploring the model's performance in more challenging clinical scenarios. The authors suggest that the model can be implemented into automated colonoscopy report generation, photo-documentation, and quality auditing solutions, indicating a clear path towards clinical integration and further development. The exclusion of narrow-band imaging frames also presents an opportunity for future research to incorporate this advanced imaging technique.

Summary updated

Abstract

Abstract Background Identification and photo-documentation of the ileocecal valve (ICV) and appendiceal orifice (AO) confirm completeness of colonoscopy examinations. We hypothesized that an artificial intelligence (AI)-empowered solution could help us automatically differentiate anatomical landmarks such as AO and ICV from polyps and normal colon mucosa. Purpose We aimed to develop and test a deep convolutional neural network (DCNN) model that can automatically identify ICV and AO, and differentiate these landmarks from normal mucosa and colorectal polyps. Method We prospectively collected annotated full-length colonoscopy videos of 318 patients undergoing outpatient colonoscopies. We created three non-overlapping training, validation, and test datasets with 25,444 unaltered frames extracted from the colonoscopy videos showing four landmarks/image classes (AO, ICV, normal mucosa, and polyps). For each landmark, we extracted an average of 30 frames for each time of its appearance. All the extracted frames were reviewed and annotated by a team of three clinicians. Using a quality assessment tool, the clinicians examined a total of 86,754 frames (7982 AO, 8374 ICV, 32,971 polyps, and 37,427 normal mucosa) and verified whether or not the frame contained one unique landmark. For this research, all frames were extracted from the white-light colonoscopies, and all narrow-band imaging frames were excluded. A DCNN classification model was developed, validated, and tested in separate datasets of images. The primary outcome was the proportion of patients in whom the AI model could identify both ICV and AO, and differentiate them from polyps and normal mucosa, with an accuracy of detecting both AO and ICV above a threshold of 40% (representing a value in which reliable identification of the landmarks can be assumed without increasing false-positive alerts). Result(s) We trained a DCNN AI model on 21,503 unaltered frames extracted from the recorded colonoscopy videos of 272 patients, and validated and tested the model on 1,924 (25 patients) and 2,017 (21 patients) unaltered frames, respectively. We applied a transfer learning technique to fine-tune the model parameters to the endoscopic images using a cross-entropy loss function and back-propagation algorithm. After training and validation, the DCNN model could identify both AO and ICV in 18 out of 21 patients (85.71%), if accuracies were above the threshold of 40%. The accuracy of the model for differentiating AO from normal mucosa, and ICV from normal mucosa were 86.37% (95% CI 84.06% to 88.45%), and 86.44% (95% CI 84.06% to 88.59%), respectively. Furthermore, the accuracy of the model for differentiating polyps from normal mucosa was 88.57% (95% CI 86.60% to 90.33%). Conclusion(s) The model can reliably distinguish these anatomical landmarks from normal mucosa and colorectal polyps. It can be implemented into automated colonoscopy report generation, photo-documentation, and quality auditing solutions to improve colonoscopy reporting quality. Please acknowledge all funding agencies by checking the applicable boxes below Other Please indicate your source of funding; MEDTEQ Disclosure of Interest M. Taghiakbari: None Declared, S. Hamidi Ghalehjegh Employee of: Imagia Canexia Health Inc. , E. Jehanno Employee of: Imagia Canexia Health Inc. , T. Berthier Employee of: Imagia Canexia Health Inc. , L. di Jorio Employee of: Imagia Canexia Health Inc. , A. N. Barkun Grant / Research support from: co-awardee in funded research projects with Imagia Canexia Health Inc., Consultant of: Medtronic Inc. and A.I. VALI Inc, E. Deslandres: None Declared, S. Bouchard: None Declared, S. Sidani: None Declared, Y. Bengio: None Declared, D. von Renteln Grant / Research support from: ERBE, Ventage, Pendopharm, and Pentax, Consultant of: Boston Scientific and Pendopharm
BibTeX
@article{taghiakbari2023automated,
  author = {Mahsa Taghiakbari and S Hamidi Ghalehjegh and E Jehanno and T Berthier and L di Jorio and Alan Barkun and Érik Deslandres and Sioui Maldonado Bouchard and Sacha Sidani and Yoshua Bengio and Daniel von Renteln},
  title = {{A108 AUTOMATED DETECTION OF ILEOCECAL VALVE, APPENDICEAL ORIFICE, AND POLYP DURING COLONOSCOPY USING A DEEP LEARNING MODEL}},
  journal = {Journal of the Canadian Association of Gastroenterology},
  year = {2023},
  doi = {10.1093/jcag/gwac036.108},
}
Figure