Introduction

Device-assisted enteroscopy (DAE) allows deep exploration of the gastrointestinal (GI) tract, combining its diagnostic abilities with tissue sampling and the application of endoscopic therapy. Although the application of artificial intelligence (AI) is growing exponentially in various imaged-based gastroenterology procedures, there is still a lack of evidence of the AI technical feasibility and clinical applicability of DAE. Convolutional Neural Networks (CNN) are a multi-layer AI architecture with high performance levels for image analysis. This study aimed to develop and test a multi-brand CNN-based algorithm for automatically detecting protruding lesions in DAE images.

Aims & Methods

​​​​A unicentric retrospective study was conducted for the development of a CNN, based on a total of 258 DAE exams. DAE was performed by two experienced endoscopists using the double-balloon enteroscopy system Fujifilm EN-580T (n = 152), the single-balloon enteroscopy system Olympus EVIS EXERA II SIF-Q180 (n = 98) and the Olympus PowerSpiral Motorized Enteroscope PSF-1 (n = 8). A total of 32043 images from small bowel were included, 3162 images containing protruding lesions (polyps, epithelial tumors, subepithelial lesions and nodules). The remaining images showed normal mucosa. A training dataset comprising 80% of the total pool of images was used for development of the network. The output provided by the network was compared to a consensus classification provided. Our primary outcome measures were sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and an area under the curve precision–recall curve (AUC-PR).

Results

The CNN model for automatic detection of GI protruding lesions identified these lesions with a sensitivity of 92.9%, a specificity of 99.0%, positive and negative predictive values of 89.6% and 90.2%, respectively. The algorithm’s accuracy was 98.4%. The AUC-PR was 1.00. The CNN analyzed images at a rate of approximately 260 frames per second, enabling AI live application in a real-life clinical setting in DAE.

Conclusion

The authors have developed a pioneer combined deep learning system for automatic detection of protruding lesions in different segments of the gastrointestinal tract. To the best of our knowledge, this is the first study regarding the automatic multi-brand panendoscopic detection using Olympus PowerSpiral Motorized Enteroscope, overcoming a relevant interoperability challenge. Our results highlight that using a CNN to detect protruding lesions is associated with high overall accuracy. The development of these tools may enhance the diagnostic yield of device-assisted enteroscopy in the setting of AI application in digestive endoscopy, particularly for panendoscopic evaluation.

References

- Rondonotti E, Koulaouzidis A, Yung DE, et al. Neoplastic diseases of the small bowel. Gastrointest Endosc Clin N Am 2017;27:93–112
- Pennazio M, Spada C, Eliakim R, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: ESGE clinical guideline. Endoscopy 2015;47:352–76.
- Rondonotti E, Spada C, Adler S, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: ESGE technical review. Endoscopy 2018;50:423–46.
- Mascarenhas, M.; Afonso, J.; Ribeiro, T.; Andrade, P.; Cardoso, H.; Macedo, G. The Promise of Artificial Intelligence in Digestive Healthcare and the Bioethics Challenges It Presents. Medicina 2023, 59, 790. https://doi.org/10.3390/medicina59040790
- Cardoso, Pedro MD, MSc1,2,*; Saraiva, Miguel Mascarenhas MD, MSc1,2,3,*; Afonso, Ribeiro, Tiago MD, MSc; Andrade, Patrícia MD; Ferreira, João PhD4,5; Cardoso, Hélder MD, MSc1,2,3; Macedo, Guilherme PhD. Artificial Intelligence and Device-Assisted Enteroscopy: Automatic Detection of Enteric Protruding Lesions Using a Convolutional Neural Network. Clinical and Translational Gastroenterology 13(8):p e00514, August 2022. | DOI: 10.14309/ctg.0000000000000514
- Martins, M.; Mascarenhas, M.; Afonso, J.; Ribeiro, T.; Cardoso, P.; Mendes, F.; Cardoso, H.; Andrade, P.; Macedo, G. Deep-Learning and Device-Assisted Enteroscopy: Automatic Panendoscopic Detection of Ulcers and Erosions. Medicina 2023, 59, 172. https://doi.org/10.3390/medicina59010172

Introduction

Capsule Endoscopy (CE) is a minimally invasive technique that allows for detailed visualization of the small bowel mucosa. Anemia and gastrointestinal (GI) bleeding are among the primary indications for CE. Accurate detection of vascular lesions is critical in this context; however, CE interpretation is time-consuming and subject to human error. Artificial intelligence (AI) has the potential to assist in this task, yet real-world studies assessing its clinical impact remain limited.

Aims & Methods

This study aimed to evaluate the performance of a deep learning convolutional neural network in detecting and identifying vascular lesions—specifically red spots and angiectasias—in complete CE videos.
We conducted a multicenter, prospective validation study comparing AI-assisted CE reading with standard, conventional interpretation. A total of 211 CE videos, acquired using three different devices across seven centers in Portugal, Spain, Brazil, and the United States, were included. Following initial standard readings, an independent expert performed AI-assisted readings using a deep learning model designed to detect and classify small bowel vascular lesions. Both reports were reviewed by a third-party expert from an independent center who adjudicated discrepant cases to establish the reference standard. Diagnostic performance was evaluated using standard metrics and the area under the receiver operating characteristic curve (AUC-ROC).

Results

AI-assisted reading identified 241 confirmed vascular lesions, compared to 158 detected through conventional reading. Sensitivity was significantly higher for AI-assisted reading (96.4% vs. 67.7%), though specificity was lower (75.7% vs. 93.5%). Overall accuracy was also higher with AI (89.7% vs. 76.1%). The AUC-ROC was 0.93 for AI-assisted and 0.78 for standard reading. Moreover, AI-assisted reading reduced the mean examination time to 316 seconds per video.

Conclusion

AI-assisted interpretation of full-length CE videos significantly outperforms conventional reading in the detection and classification of vascular lesions. The inclusion of videos from multiple CE systems addresses interoperability challenges, and the involvement of patients from four countries across two continents supports the model's generalizability and real-world clinical applicability.

Introduction

While capsule endoscopy (CE) offers a minimally invasive way to examine the entire digestive tract, current analysis methods are time-consuming and susceptible to errors. Protruding lesions are frequentlyencountered with varying significance throughout the GI tract. This study aims to develop and test a Convolutional Neural Network for panendoscopic automatic detection of protruding lesions on CE exams.

Aims & Methods

A multicentric retrospective study was performed on 1222 CE exams. A total of 191455 images from six different CE device types were evaluated, with 52717 confirmed protruding lesions (epithelial andsubepithelial lesions) after a rigorous triple-validation process. First, we conducted a 5-fold cross validation with exam-split for assessing CNN’s robustness. Then, we evaluated CNN’s global performance in an exam-split train and test division of data, assessing the latter. Primary outcomes measures were sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV) and area under the conventional receiver operating characteristic curve (AUC-ROC) and the precision-recall curve (AUC-PR).

Results

Regarding test set, the sensitivity were 90.7% and specificity 94.0%. The PPV and NPV were 95.3% and 88.3%, respectively. The global accuracy was 92.1%. The CNN processed 112 frames per second.

Conclusion

This study aims to address the gap in AI-enhanced capsule panendoscopy by investigating the development of the first CNN for the detection of protruding lesions across the GI tract. AI's improvement of CE's diagnostic accuracy, along with the growing interest in minimally invasive procedures, may contribute to increasing access to this diagnostic tool.

Introduction

Capsule Endoscopy (CE) offers a minimally invasive method for comprehensive visualization of the gastrointestinal (GI) tract. Protruding lesions are relatively common findings in CE, with variable incidence and diverse pathological significance. Accurate detection of these lesions is crucial but remains challenging due to the time-consuming nature of CE interpretation and susceptibility to human error. While artificial intelligence (AI) presents a promising solution, real-world evidence evaluating its clinical performance is still scarce. To assess the diagnostic performance of a deep learning convolutional neural network (CNN) for the detection of protruding lesions in complete CE videos under real-world clinical conditions.

Aims & Methods

Our group conducted a prospective, multicenter validation study comparing AI-assisted CE reading to standard, conventional interpretation. A total of 211 CE videos, obtained using three different CE systems across seven centers in Portugal, Spain, Brazil, and the United States, were analyzed. Following initial standard readings, an independent expert conducted AI-assisted readings using a CNN specifically trained to detect small bowel protruding lesions. Discrepancies between conventional and AI-assisted reports were adjudicated by a third-party expert from an independent center to establish the reference standard. Diagnostic performance was assessed using sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve (AUC-ROC).

Results

AI-assisted reading detected 106 confirmed protruding lesions, compared to 42 identified through conventional reading. Sensitivity was significantly higher with AI-assisted reading (95.9% vs. 64.6%), although specificity was lower (78.7% vs. 96.1%). Overall accuracy was relatively lower with AI-assisted reading (81.2% vs. 91.5%). The AUC-ROC was 0.86 for AI-assisted and 0.80 for standard reading. Additionally, AI-assisted reading reduced the mean examination time to 316 seconds per video.

Conclusion

AI-assisted interpretation of complete CE videos significantly enhances the detection of protruding lesions compared to conventional reading, although at the expense of slightly reduced specificity. The inclusion of videos from multiple CE systems addresses interoperability challenges, while the participation of centers from four countries across two continents supports the model’s generalizability and real-world clinical applicability.

Introduction

Device-assisted enteroscopy (DAE) allows deep exploration of the gastrointestinal (GI) tract, combining its diagnostic abilities with tissue sampling and the application of endoscopic therapy. Although the application of artificial intelligence (AI) is growing exponentially in various imaged-based gastroenterology procedures, there is still a lack of evidence of the AI's technical feasibility and clinical applicability of DAE. Convolutional Neural Networks (CNN) are a multi-layer AI architecture with high-performance levels for image analysis.

Aims & Methods

This study aimed to develop and test a multi-brand CNN-based algorithm for automatically detecting multiple lesions in DAE.
A unicentric retrospective study was conducted to develop a CNN based on 258 DAE exams. DAE was performed by two experienced endoscopists using the double-balloon enteroscopy system Fujifilm EN-580T (n = 152), the single-balloon enteroscopy system Olympus EVIS EXERA II SIF-Q180 (n = 98), and the Olympus PowerSpiral Motorized Enteroscope PSF-1 (n = 8). A total of 11205 images from different segments of the GI tract (stomach, small bowel, and colon) were included, 2535 images containing protruding lesions (polyps, epithelial tumors, subepithelial lesions, and nodules), 1435 containing blood or hematic residues, 1395 images containing angioectasias and 450 images containing ulcers/erosions. The remaining images showed normal mucosa (n=5390). A training dataset comprising 80% of the total pool of images (n=8964) was used to develop the network. The output provided by the network was compared to a consensus classification provided. Our primary outcome measures were sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and an area under the curve precision–recall curve (AUC-PR).

Results

The CNN model for automatic detection of GI protruding lesions identified these lesions with a sensitivity of 97.0%, a specificity of 97.4%, and positive and negative predictive values of 94.6% and 98.6%, respectively. The component for automatic detection of blood/hematic residues had a sensitivity of 95.8%, a specificity of 97.6%, and positive and negative predictive values of 91.4% and 98.9%, respectively. The CNN’s angioectasia model automatically detected angioectasia with a sensitivity of 88.5%, a specificity of 97.1%, and positive and negative predictive values of 88.1% and 97.0%, respectively. Furthermore, the CNN ulcers/erosions model automatically detected ulcers/erosions with a sensitivity of 100%, a specificity of 96.4%, and positive and negative predictive values of 96.8% and 100%, respectively. The algorithm’s accuracy was 98.7%. The AUC-PR was 1.00. The CNN analyzed images at approximately 239 frames per second, enabling AI live application in a real-life clinical setting in DAE.

Conclusion

The authors have developed a pioneer combined deep learning system for automatically detecting a wide range of lesions in different gastrointestinal tract segments. To our knowledge, this is the first study regarding the automatic multi-brand panendoscopic detection using Olympus PowerSpiral Motorized Enteroscope, overcoming a relevant interoperability challenge. Our results highlight that using a CNN to detect pleomorphic lesions is associated with high overall accuracy. The development of these tools may enhance the diagnostic yield of device-assisted enteroscopy in the setting of AI applications in digestive endoscopy, particularly for panendoscopic evaluation.

References

1. Goenka, M.K.; Majumder, S.; Kumar, S.; Sethy, P.K.; Goenka, U. Single center experience of capsule endoscopy in patients with obscure gastrointestinal bleeding. World J. Gastroenterol. 2011, 17, 774–778. [CrossRef] [PubMed]
2. Shahidi, N.C.; Ou, G.; Svarta, S.; Law, J.K.; Kwok, R.; Tong, J.; Lam, E.C.; Enns, R. Factors associated with positive findings from capsule endoscopy in patients with obscure gastrointestinal bleeding. Clin. Gastroenterol. Hepatol. 2012, 10, 1381–1385. [CrossRef] [PubMed]
3. Iddan, G.; Meron, G.; Glukhovsky, A.; Swain, P. Wireless capsule endoscopy. Nature 2000, 405, 417. [CrossRef]
4. Aoki, T.; Yamada, A.; Aoyama, K.; Saito, H.; Tsuboi, A.; Nakada, A.; Niikura, R.; Fujishiro, M.; Oka, S.; Ishihara, S.; et al. Automatic detection of erosions and ulcerations in wireless capsule endoscopy images based on a deep convolutional neural network. Gastrointest. Endosc. 2019, 89, 357–363.e2. [CrossRef] [PubMed]
5. Pennazio, M.; Rondonotti, E.; Despott, E.J.; Dray, X.; Keuchel, M.; Moreels, T.; Sanders, D.S.; Spada, C.; Carretero, C.; Cortegoso Valdivia, P.; et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Guideline—Update 2022. Endoscopy 2023, 55, 58–95. [CrossRef]

Introduction

Device-assisted enteroscopy (DAE) offers comprehensive examination of the gastrointestinal tract, but its efficacy is dependent on adequate bowel preparation. While prior research explored AI models for colonoscopy cleansing, there is a lack of studies evaluating both small bowel and colon cleansing during DAE.

Aims & Methods

The aim of this study was to develop and validate a convolutional neural network (CNN) capable of automatically classifying bowel cleansing levels during DAE, encompassing both small bowel and colon regions. We retrospectively analyzed 28 procedures (single balloon, double-balloon and motorized spiral enteroscopy) from a high-volume center. Bowel preparation was graded as excellent (≥90% visible mucosa), satisfactory (50-90%), or unsatisfactory (<50%). A dataset of 88,623 images (training: 90%, testing: 10%) was used, covering both small bowel and colon areas. The CNN's performance was evaluated against expert consensus using sensitivity, specificity, accuracy, and AUC-ROC.

Results

We included 51948 frames that were considered excellent, 28820 frames that were considered satisfactory and 7855 frames that were considered unsatisfactory. The CNN demonstrated the following performance metrics: excellent cleansing (sensitivity: 97.8%, specificity: 80.3%, accuracy: 90.6%, AUC-ROC: 0.95), satisfactory cleansing (sensitivity: 81.8%, specificity: 97.9%, accuracy: 92.7%, AUC-ROC: 0.95), and unsatisfactory cleansing (sensitivity: 68.7%, specificity: 99.5%, accuracy: 96.8%, AUC-ROC: 0.96).

Conclusion

This study presents the first globally developed CNN for accurate classification of bowel cleansing throughout DAE, encompassing both small bowel and colon. This advancement represents a necessary step for optimizing diagnostic accuracy and therapeutic efficacy in gastrointestinal procedures, namely DAE, underscoring the pivotal role of innovative technologies in advancing medical practice