Capsule Endoscopy in Inflammatory Bowel Disease: Panenteric Capsule Endoscopy and Application of Artificial Intelligence

Abstract

Video capsule endoscopy (VCE) of the small-bowel has been proven to accurately diagnose small-bowel inflammation and to predict future clinical flares among patients with Crohn’s disease (CD). In 2017, the panenteric capsule (PillCam Crohn’s system) was introduced for the first time, enabling a reliable evaluation of the whole small and large intestines. The great advantage of visualization of both parts of the gastrointestinal tract in a feasible and single procedure, holds a significant promise for patients with CD, enabling determination of the disease extent and severity, and potentially optimize disease management. In recent years, applications of machine learning, for VCE have been well studied, demonstrating impressive performance and high accuracy for the detection of various gastrointestinal pathologies, among them inflammatory bowel disease lesions. The use of artificial neural network models has been proven to accurately detect/classify and grade CD lesions, and shorten the VCE reading time, resulting in a less tedious process with a potential to minimize missed diagnosis and better predict clinical outcomes. Nevertheless, prospective, and real-world studies are essential to precisely examine artificial intelligence applications in real-life inflammatory bowel disease practice.

INTRODUCTION

Video capsule endoscopy (VCE) is a noninvasive modality for visualizing the mucosal surface of the small and large intestines.¹ VCE of the small-bowel has been proven to precisely diagnose small-bowel inflammation and predict future clinical flares among patients with Crohn’s disease (CD).^2-6 Small-bowel VCE is a prime modality for diagnosis and monitoring of patients with CD, and consequently, to prevent disease progression and complications (i.e., intestinal-surgery, clinical exacerbation) among this population.^7-9

While small-bowel VCE had been widely and beneficially used among patients with CD much earlier,¹⁰ colon capsule endoscopy use in this population has been first reported in 2014.¹¹ Thenceforth several studies have been published describing colon capsule endoscopy performance in patients with CD.^8,11-14 Though initially colon capsule endoscopy has been claimed to underestimate colonic lesions compared with optical colonoscopy,¹² it was considered as a promise due to the higher rates of terminal ileum lesion detection compared with the traditional procedure,^8,11,13,14 covering an extended area of the gastrointestinal tract. Emboldened by this advantage, a novel panenteric capsule had been developed and introduced in 2017¹⁵–the PillCam Crohn’s system (PCC; Medtronic, Yokneam, Israel), in which the capsule and its software were tailored to patients with CD,¹⁶ allowing visualization of both the small and large intestines.

Machine-learning technology is a subclass of artificial intelligence (AI), affecting many aspects of medical practice,¹⁷ including several medical fields such as radiology, dermatology, gastroenterology and ophthalmology.¹⁸ Deep learning is a subclass of machine-learning which is mainly based on artificial neural networks.^17,18 In the recent years, applications of deep-learning, including convolutional neural networks (CNN), for VCE have been well studied, demonstrating an accurate performance for detection of various gastrointestinal pathologies (e.g., gastrointestinal bleeding,^19-22 angioectasias,^23-26 esophagus and small-bowel mucosal ulcers^26-31).

In this paper we aimed to review the accumulating data regarding the use of PCC among patients with CD. We also aimed to describe the innovative and emerging use of AI among patients with CD undergoing VCE.

PANENTERIC CAPSULE IN PATIENTS WITH CD

In 2017, a new panenteric VCE was introduced.¹⁶ The PCC is a two-headed capsule with a field of view of 344°, along with an adaptive frame rate technology which obtains up to 35 frames per second adapting to the speed of transit, allowing better tissue coverage and battery conservation.¹⁶ The PCC system platform and software have a novel assessment for inflammatory disease, specifically CD. The software divides the small bowel into three anatomic segments according to their length, as well as the colon. Three key assessment parameters are then assessed: disease distribution, lesion severity and linear extent. The most severe lesion and most common lesion in each segment are documented (Fig. 1).

Fig. 1.

(A) PillCam Crohn's capsule, DR3 data recorder and wireless sensors. (B) A representative graphic of a patient with active Montreal L1 and images of small bowel lesions. (C) RAPID^TM Reader Software breaks down small bowel segment based on identified anatomical landmarks. The reader classifies the most severe and most common lesion (none, mild, moderate and severe), presence or absence of stricture and extent of disease (0%–10%, 10%–30%, 30%–60%, 60%–100% of segment). Adapted from Eliakim R, et al. Endosc Int Open 2018;6:E1235-E1246¹⁶ and Tai FW, et al. United European Gastroenterol J 2021;9:248-255.³³

CLINICAL PRACTICE USING THE PCC

Leighton et al.¹⁵ demonstrated an improved performance of the PCC compared with ileo-colonoscopy (IC) among 66 CD patients with active disease who underwent both procedures. Either per-subject diagnostic yield rate or per-segment diagnostic yield rate for active CD lesions, was higher in the PCC compared with the IC procedure (83.3% vs 69.7% [yield difference, 13.6%; 95% confidence interval, 2.6% to 24.7%], and 40.6% vs 32.7% [yield difference, 7.9%; 95% confidence interval, 3.3% to 12.4%], respectively). Considering the substantial rate of active lesions detected only by the PCC (18%, 12/66), in which only one was limited to the proximal bowel, it was concluded that this procedure should be at least a complementary one to the IC among patients with CD. Leighton’s study used the capsule without its specific software. In 2020, a study performed by Bruining et al.,³² included 99 patients with established non-stricturing CD to assess the performance of PCC compared with IC and/or magnetic resonance enterography (MRE), in the detection of mucosal lesions in this population. The authors demonstrated comparable sensitivity rates and higher specificity rates in the overall intestinal assessment between the PCC compared to the MRE and/or IC (94% vs 100%, p=0.125 and 74% vs 22%, p=0.001, respectively). PCC had higher sensitivity and specificity in the proximal small bowel compared with MRE, higher specificity in the terminal ileum compared with MRE and/or IC, and equal performance in the colon compared to the IC. Patients’ satisfaction was superior for the capsule compared with the two other procedures. These findings emphasized the great advantage of PCC to enable a reliable disease staging of anatomic involvement among patients with CD, while undergoing a single procedure. Tai et al.³³ examined the PCC performance in predicting the need of treatment intensification among 93 patients with CD (22 suspected CD, 71 established CD). PCC detected active disease in 48 out of 71 (67.6%) patients with established CD, and in three out of 22 patients (13.6%) with suspected CD. Disease extent was upstaged in 24 out of 71 (33%) patients with CD, of them, nine patients with newly upper gastrointestinal tract involvement. Overall, PCC findings led to treatment intensification in 36 out of 93 (39%) patients, and it was associated with proximal small-bowel involvement. Neither symptoms nor biochemical markers (i.e., fecal calprotectin, C-reactive protein) reliably identified active CD compared with PCC. This study demonstrated the important role of PCC in diagnosing, disease staging and optimizing disease treatment among patients with suspected or established CD. Oliva et al.,³⁵ assessed the yield of PCC among 48 pediatric patients with quiescent CD (Crohn’s Disease Activity Index³⁴ <10) to monitor mucosal healing and deep remission in a treat-to-target strategy. The PCC detected significant inflammation in 34 out of 48 (71%) patients involving the small-bowel or the colon (16/26 patients in clinical remission vs 18/22 patients with clinical activity). Accordingly, these findings led to treatment change in 34 out of 48 (71%) patients at baseline and in 11 out of 48 (23%) patients at 24 weeks follow-up. As a result, mucosal healing rate increased from 21% at baseline to 58% at week 52. Thus, PCC treat-to-target approach, led to higher rates of mucosal healing and deep remission among this population. In 2020, Eliakim et al.,³⁶ evaluated the accuracy of a novel scoring system for PCC including 41 patients with CD. For each small bowel tertile, the Lewis score (LS) was extracted using the automated calculator embedded in the software. Similarly, LS was calculated for the right and left colon. The small-bowel LS was derived of the score of the tertile with the most significant disease involvement plus stricture score. The correlation of Eliakim score for PCC to fecal calprotectin was higher than between LS and fecal calprotectin (r=0.32 and r=0.54 respectively, p=0.001 for both). Table 1 summarizes PCC studies as mentioned above.

Table 1.

Summary of the PillCam Crohn’s Capsule (PCC) Studies

Study (year)	Study design	Patients	Comparative procedure	Performance measures	Safety
Leighton et al. (2017)15	Prospective	66 Patients with active CD	IC	83.3% and 69.7% of CD lesions in the PCC group vs IC group, respectively	1: Obstructive symptoms following PCC procedure
					1: GIT symptoms following PC ingestion
					1: GIT symptoms following bowel preparation protocol
Eliakim et al. (2018)16	Prospective feasibility study	41 Patients with established or suspected IBD	-	-	No retained capsule was reported
Bruining et al. (2020)32	Prospective	99 Patients with non-stricturing CD	IC, MRE	Comparable sensitivity rate between PCC and either IC or MRE	1: Partial bowel obstruction due to retained capsule
				Higher specificity rate compared with MRE in detection of small-bowel lesions. Comparable specificity compared to IC in detection of TI and colon lesions	1: Sigmoid perforation during IC
Tai et al. (2021)33	Observational	93 Patients (22 suspected CD, 71 established CD)	-	33%: Upstaging of disease extent	2: Retained capsule (small-bowel and colon strictures)
				9 Patients with newly upper GIT involvement
				39%: Disease management change
Oliva et al. (2018)35	Prospective	48 Pediatric patients with quiescent CD	-	Disease management change in 71% and 23% at baseline and at 24 wk, respectively	3: Nausea and vomiting following bowel preparation protocol
				Accordingly, mucosal healing rate increased from 21% (baseline) to 58% (52 wk)
Eliakim et al. (2020)36	RCT	41 Patients with CD	Lewis score, fecal calprotectin	Better correlation of Eliakim score to fecal calprotectin than Lewis score to fecal calprotectin	No capsule retention

CD, Crohn’s disease; IC, ileo-colonoscopy; GIT, gastrointestinal tract; PC, patency capsule; IBD, inflammatory bowel disease; MRE, magnetic resonance enterography; TI, terminal ileum; RCT, randomized control trial.

BOWEL PREPARATION AND INTESTINAL CLEANSING

Table 2 summarizes bowel preparation protocols and cleaning performance in PCC studies. All the studies’ protocols used polyethylene glycol based solution prior to the capsule ingestion (two divided doses of 1.5 to 2 L administered in the evening before and in the morning of the examination day). Food but a clear liquid diet was prohibited on the day before and the examination day. Additional laxative was used upon the capsule had been reached to the small bowel. Bowel cleansing was graded as poor, fair, good or excellent.³⁷ Comparing bowel cleansing level between PCC and IC, there was no difference in small-bowel cleansing, while colon cleansing was significantly better in the latter procedure.^15,32 Overall bowel cleansing was better for small-bowel portions than the colon portions (good/excellent rate of 80% to 90% and up to 75%, respectively). Out of 386 patients undergoing PCC, there were only two cases in which colon preparation was inadequate to preclude reading of the colon frames.^33,36

Table 2.

Bowel Preparation Protocols and Cleaning Performance in PillCam Crohn’s Capsule (PCC) Studies

Study	Bowel preparation protocol	Preparation performance37
Leighton et al. (2017)15	Day before (–1): clear liquid diet and 2 L PEG at the evening. Day 0: 2 L PEG at the morning Optional: 10 mg Metoclopramide 1 hr after capsule ingestion. Upon small-bowel detection and (again) 3 hr later: 88 mL Suprep+1 L water 10 mg bisacodyl suppository 2 hr later	% Excellent/good cleansing level: >80% in each part of the colon and the TI for the IC procedure. For the PCC procedure: TI 87.9%, colon segments 39.4%–62.9%
Eliakim et al. (2018)16	Day (–1): clear liquid diet and 2L PEG/Fortrans/Solucion Bohm at the evening Day 0: 2 L PEG/Fortrans/Solucion Bohm at the morning Optional: 10 mg Metoclopramide 1hr after capsule ingestion (if the capsule was detected in the stomach) Upon small-bowel detection 88 mL Suprep +240 mL water or 1 sachet of PICO-SALAX and again 3 hr later 10 mg bisacodyl suppository 2 hr later	% Excellent/good cleansing level: >97.5% in the small bowel vs up to 75 % in the colon
Bruining et al. (2020)32	Day (–1): clear liquid diet 13–15 hr and 1–3 hr before PCC: 2 L PEG 10 mg Metoclopramide or 250 mg erythromycin 1 hr after PCC ingestion Upon small-bowel detection and 3 hr later: 88 mL of sodium/Magnesium sulfate diluted to 1 L of water 10 mg bisacodyl suppository 2 hr later	% Excellent/good: 79% proximal small bowel 90% TI 64% colon
Tai et al. (2021)33	Day before (–1): clear liquid diet and 2 L PEG at the evening. Day 0: 2 L PEG at the morning Optional: 10 mg Metoclopramide 1 hr after capsule ingestion. Upon small-bowel detection and (again) 3 hr later: 88 mL Suprep+1 L water 10 mg bisacodyl suppository 2 hr later	Inadequate bowel preparation: 1/93 (1.1%)
Oliva et al. (2018)35	Day (–1): clear liquid diet with 50 mL up to 2 L of PEG solution Day 0: 1 hr before capsule ingestion 50 mL up to 2 L of PEG solution Optional: 0.25 mg/kg of domperidone (for delayed gastric transit >1 hr) Once, the PCC had been detected in the small bowel - first booster of sodium sulfate 30 mL was given and another booster of 15 mL, 3 hr later Optional: 10 mg bisacodyl suppository 3.5 hr later	Small-bowel and colon cleansing level: Excellent: 30% Good: 58% Fair: 10% Poor: 2%
Eliakim et al. (2020)36	Day (–1): clear liquid diet and 1.5 L of PEG on the evening before the examination Day 0: 1.5 L an hour prior to capsule ingestion PICO-SALAX (10 mg sodium picosulfate) diluted in 75 mL of water upon small-bowel detection A 10 mg of metoclopramide PO if the capsule remained in the stomach for more than an hour after ingestion (optional) If the PCC was not excreted within 3 hr of ingestion, a second sachet of PICO-SALAX was administered 3 hr after the first one A 10 mg Bisacodyl suppository 2 hr later if the capsule was not excreted	% Excellent/good: Small bowel: 94.4% Colon: 75% % Fair and poor for the colon 22.5% and 2.5%, respectively

PEG, polyethylene glycol; TI, terminal ileum; IC, ileo-colonoscopy; PO, per os.

SAFETY PROFILE USING THE PCC AND PROCEDURE COMPLETION

Of 386 patients undergoing PCC,^{15,16,32,33,35,36} only 12 patients experienced serious adverse events (including three cases of capsule retention [<1%]^32,33) (Table 1). One case occurred despite patency confirmation by patency capsule (PC) procedure,³³ one case after MRE assurance (without PC procedure),³² and a sole case of capsule retention due to a colonic stricture,³³ though PC had passed uneventfully. Of them, one patient was hospitalized due to partial bowel obstruction.³² All cases were attributed to bowel strictures, and required PCC retrieval and stricture dilation by endoscopic procedure.^32,33 No case required surgical treatment.

Other serious adverse events included two cases of obstructive symptoms and signs after PCC and PC procedures,¹⁵ gastrointestinal tract symptoms following bowel-cleansing preparation protocol¹⁵ and sigmoid perforation during IC procedure.³² Other non-serious adverse events including nausea, vomiting and abdominal pain occurred in less than 15% of the procedures. No serious adverse events related to the PCC were reported in pediatric CD patients (Table 1), while two episodes of nausea and one episode of vomiting were reported in three patients following adherence to bowel-preparation protocol.³⁵

Tai et al.³³ reported there were eight patients (8.6%) with incomplete colon examination (excluding a colonic stricture), five were due to loss of battery power, two due to loss of capsule signal (2.2%) and one due to inadequate bowel preparation (1.1%). Oliva et al.³⁵ noted 17 out of 142 (~1.2%) procedures in which the capsule were not excreted before the battery expired, though seven of them reached the rectum, enabling a complete evaluation of the colon.

AI-BASED DETECTION OF CD LESIONS AMONG PATIENTS UNDERGOING SMALL BOWEL VCE

A single VCE procedure, captures and broadcasts an average of 12,000 images per-patient, making it tedious for a single reader reading and interpreting an entire examination. The latter may require 30 to 40 minutes on average, even for experienced VCE readers.^38,39 Leenhardt et al.⁴⁰ observed more than 80% of interobserver agreement rate in the identifying of ulcerative and inflammatory lesions during VCE reading, but still, there was a substantial rate of disagreement in VCE interpretation. Considering the monotone manner of VCE reading, with other technical challenges including no way to direct or focus the camera and the existence of only few frames for each lesion, its considerable rate of missed lesions (10%) is conceivable.⁴¹ AI-based VCE reading, including CNN algorithms to interpret VCE frames, has the potential to minimize the above-mentioned drawbacks, performing automated image analysis and interpretation.¹⁸ The CNN automatically extracts the features from raw input data (i.e., VCE frames), to identify distinct patterns in the dataset (e.g., small-bowel ulcers). The main dataset is randomly distributed into training, validation, and testing sets. The training set is used to fit the model and hyper-parameters, while the validation set is used to evaluate model performance. The testing set, which is sometimes an external dataset, provides an unbiased evaluation of the final model (Figs 2 and 3).^42,43

Fig. 2.

Visualization of dataset splits performing a convolutional neural network model.

Fig. 3.

Convolutional neural networks architecture representation.

Conv, convolutional; ReLU, rectified linear unit.

AI PERFORMANCE IN THE DETECTION OF ULCERS AND EROSIONS

Since 2018, data regarding the detection of ulcers and/or erosions using deep-learning application in VCE have been accumulated (Table 3 summarizes their main characteristics and findings).

Table 3.

Summary of the Published Studies on Artificial Intelligence for Detection of Ulcers/Erosions in the Small Bowel

Study	Study design	Algorithm	No. of patients	Cohort details	Type of lesion (No. of normal/pathological frames)	No. of frames (training/validation datasets)	Performance measures
Fan et al. (2018)31	Retrospective	AlexNet CNN	144	NA	Small-bowel ulcers and erosions (13,000/8,160)	12,910/8,250	Accuracy of 95.16% and 95.34% for the detection of ulcers and erosions, respectively
Aoki et al. (2019)30	Retrospective	CNN-based on SSD	180	Patients with various causes of erosions and ulcers*	Small-bowel ulcers and erosions (10,000/5,800)	5,360/10,440	Accuracy of 90.8% for the detection of ulcers and erosions
Wang et al. (2019)27	Retrospective	Modified	1,504	NA	Small-bowel ulcers (19,457/17,821)	32,919/4,359	Accuracy of 90.1% for the detection of ulcers
		RetinaNet
Ding et al. (2019)44	Retrospective	CNN-based auxiliary model	6,970	Patients with various small-bowel VCE findings	Pathological vs normal VCE frame† (NA)	158,235/113,268,334	Sensitivity and specificity of 99.73% and 100% for the detection of ulcers
Otani et al. (2020)26	Retrospective	Modified	194	Patients undergoing VCE for occult/overt GIB, tumor follow-up or IBD	Pathological vs normal VCE frames (34,437/5,526)	39,963/1,247	Accuracy of 98.6%–99.3% for the detection of ulcers and erosions, respectively
		RetinaNet
Klang et al. (2020)28	Retrospective	CNN	49	Patients with or without CD with ulcerated or normal mucosa	Small-bowel CD ulcers (10,249/7,391)	14,112/3,528	Accuracy of 95.4%–96.7% and 73.7%–98.2%, for per-lesion analysis and per-patient analysis, respectively
Barash et al. (2021)47	Retrospective	CNN	49	Patients with or without CD with ulcerated or normal mucosa	Grading of small-bowel CD ulcers (10,249/7,391)	1,242/248	Accuracy rate of 91% comparing grade 1 to grade 3 ulcerations
Klang et al. (2021)48	Retrospective	Google’s EfficientNet networks	NA	Patients with or without CD with ulcerated or normal mucosa	Small-bowel strictures (14,266/13,626)	NA	Accuracy rate of 93.5% and 78.9% for the detection of strictures, and for the classification of ulcerated versus non-ulcerated strictures, respectively
Hwang et al. (2021)45	Retrospective	VGGNet	NA	Patients undergoing small-bowel VCE	Classification of hemorrhagic and ulcerative small-bowel lesions (8,578/4,738)	7,556/5,760	Accuracy rate of 96.62%–96.83%
Mascarenhas Saraiva et al. (2021)50	Retrospective	Xception	4,319	Patients undergoing VCE with normal mucosa, or small-bowel pathology (polyps, ulcers, vascular lesions, etc.)	Classification of higher risk small-bowel lesions for bleeding (18,010/35,545)	42,844/10,711	AUC for ulcer detection of 0.99 for P1 ulcers, and 1.00 for P2 ulcers‡
Afonso et al. (2022)49	Retrospective	CNN	NA	Patients with small-bowel ulcers, erosions, or normal small-bowel VCE	Classification of potential risk for bleeding of small-bowel ulcers and erosions (1,897/4,233)	4,904/1,226	Accuracy rate of 95.6% for the detection and classification of erosions and/or ulcers (with any bleeding potential)
Majtner et al. (2021)52	Prospective	ResNet-50	38	Patients with suspected or known CD undergoing PCC	Detection of small-bowel or colon CD lesions and classification of the severity of these lesions (4,996/2,748)	5,419/767	Accuracy rate of 98.4%–98.6% to detect small-bowel and or colon inflammatory/ulcerative lesions
Ferreira et al. (2022)53	Retrospective	Xception	NA	Patients undergoing PCC	Detection of small-bowel or colon ulcers or erosions (19,190/5,300)	19,740/4,935	Accuracy rate of 98.8% for the detection of ulcers and erosions
Kratter et al. (2022)46	Retrospective	EfficientNet	NA	Patients with and without CD undergoing small-bowel VCE or PCC	Detection of small-bowel or colon ulcers or erosions (15,684/17,416)	NA	Accuracy rate of 97.4% for the detection and the grading of mucosal ulcers in different VCE types

CNN, convolutional neural network; NA, not applicable; SSD, single-shot detector; VCE, video capsule endoscopy; GIB, gastrointestinal bleeding; IBD, inflammatory bowel disease; CD, Crohn’s disease; AUC, area under the curve; PCC, PillCam Crohn’s capsule.

*Patients who used nonsteroidal anti-inflammatory drugs (26%), patients with IBD (11%), small-bowel malignancy (7%), anastomotic ulcer (6%), ischemic enteritis (2%), Meckel diverticulum (2%), radiation enteritis (1%), miscellaneous (3%), and unknown cause (45%); ^†The dataset contained frames of various small-bowel lesions including ulcers, erosions, vascular lesions, tumors, polyps, protruding lesion, vascular lesion, bleeding, parasites, and diverticulum and normal small-bowel VCE frames; ^‡Based on Saurin’s classification.⁵¹

Fan et al.³¹ presented a novel computer-aided method to detect ulcers and erosions in the small-bowel with a high accuracy rate (>95%). This model demonstrated higher sensitivity rate in the detection of ulcers compared with erosions, probably due to more distinctive features of the former compared with the latter. Though excellent performance has been achieved, there were about 5% of false positive rate. Aoki et al.,³⁰ trained a CNN system, to detect small-bowel ulcerations and erosions from a pool of frames, originated from various small-bowel pathologies (nonsteroidal anti-inflammatory drugs, inflammatory bowel disease, malignancy, etc.). Sensitivity rates were almost comparable either in nonsteroidal anti-inflammatory drugs or CD originated frames (~90%). Despite the high-speed review time by the model (44.8 images per second), it had still identified three pathological frames, which were missed by the conventional readers. Hence, emphasizing its great advantage in the detection of fine features in the frame. On the other hand, high degree of obscuration due to bubbles, debris, and bile led to 11.8% of false negative. Wang et al.²⁷ used a second glance detection framework to detect small-bowel ulcers. Their model both classified images and also provided bounding box for lesion localization. In comparison to previously studied frameworks (RetinaNet, Faster-RCNN), using the second glance improved small ulcer detection by 10%. Still, ulcer size had a prime effect on the detection rate (92% vs 85% for ulcer size >1% and <1% of the whole image, respectively).

Two previously published studies have focused on a diagnosis of multiple types of lesions (i.e., ulcers, vascular lesions, tumors, etc.) rather than a single one.^26,44 In a multicenter study, Ding et al.⁴⁴ presented a CNN-based model auxiliary with increased detection rate by 20% compared with conventional reading, either per-lesion or per-patient analysis. Interestingly, though CNN-based model was significantly more sensitive in the detection of small-bowel ulcers, sensitivity has been improved only by 1% to 2% compared with conventional reading. Otani et al.,²⁶ used a modified RetinaNet CNN, demonstrating an accuracy rate of ~99% in the detection of small-bowel ulcers and erosions as well as for vascular lesions. Notably, the modified RetinaNet CNN model had a higher area under the curve (AUC) value compared with the Single-Shot MultiBox Detector based AI system, which had predominantly been used in previously published studies. The latter’s performance has been proven among external validation cohort from 40 patients, though with a more modest AUC value (0.996 vs 0.928). Ding et al.⁴⁴ and Otani et al.²⁶ reported four and three cases of missed pathological diagnosis, respectively.

Klang et al.,²⁸ retrospectively collected VCE frames from 49 patients with and without CD, to evaluate CNN performance in the detection of small-bowel ulcers. Like similar studies, the authors noted an impressive accuracy rate for the detection of ulcers using the trained dataset. However, the model’s accuracy rate on unseen patients ranged from 73.7% to 98.2%, probably reflecting real-world practice. The average duration for detecting a complete film was 204.7±93.9 seconds.

Hwang et al.⁴⁵ trained CNN model in two different ways: a combined model (hemorrhagic and ulcerative lesions trained separately) and a binary model (all abnormal images trained without discrimination). Both the combined and binary models acquired high accuracy for lesion detection, and the difference between the two models was not significant (96.83% vs 96.62%, p=0.122). However, there were higher sensitivity and negative predictive value rates of the combined model compared with the binary one, leading to lower rates of missed diagnosis (23 cases vs 47 cases).

Recently, Kratter et al.⁴⁶ developed a combined model for two different capsules (PillCam Crohn and PillCam SB3, Medtronic), with excellent performance in detection of intestinal ulcers (accuracy rate of 97.4%), providing an essential tool in real-life practice of patients with CD in which several types of VCE may be used.

ULCER SEVERITY GRADING, DETECTION OF STRICTURES AND BLEEDING POTENTIAL ASSESSMENT

As part of inflammatory lesion detection (i.e., ulcer and erosion), several studies have been focused on classifying lesions based on its distinct parameters, to better predict disease course and personalize disease management.

In 2021, Barash et al.⁴⁷ demonstrated a novel use of ordinal CNN model for ulcer severity grading among patients with CD. Severity grading of CD ulcers was based on the PillCam CD classification (grade 1-3 from mild to severe) (Fig. 4). The best performance was in distinguishing between grade 1 to grade 3 ulcerations, achieving an accuracy rate of 91%. In differentiating between grade 2 to either grade 1 or grade 3, the performance was less impressive (accuracy of 65% and 79%, respectively), consistent with the performance of the conventional reading method (distinction of severity category, involving grade-2 ulcer achieved up to 40% of accuracy rate). Kratter et al.,⁴⁶ demonstrated similar results in classification ulcerations to grade 1 and grade 3, with AUC of 0.99.

Fig. 4.

Severity grading of small-bowel ulcers (A: mild, B: moderate, C: severe) based on the PillCam Crohn's disease classification. Adapted from Barash Y, et al. Gastrointest Endosc 2021;93:187-192, with permission from Elsevier.⁴⁷

Klang et al.⁴⁸ evaluated the ability of a neural network model in identifying intestinal strictures for the first time. For classifying stricture versus non-stricture lesions, the network exhibited an average accuracy of 93.5%. The ulcerated versus non-ulcerated strictures classification network resulted in an accuracy rate of 78.9% (Fig. 5).

Fig. 5.

Class activation maps (heatmaps) of an ulcer image. Heatmaps enabled a visual presentation of image regions which led to lesion classification. Adapted from Klang E, et al. J Crohns Colitis 2021;15:749-756, with permission from Oxford University Press.⁴⁸

Two studies from the same group demonstrated a novel CNN model to identify and classify small-bowel ulcers (among other enteric lesions) in whom having high risk for bleeding^49,50 based on Saurin classification.⁵¹ According to Saurin classification, the hemorrhagic potential of ulcers was depended on their size: small ulcers were regarded as P1 lesions, while large ulcerations (>20 mm) were regarded as P2 lesions. Mascarenhas Saraiva et al.⁵⁰ showed that among a wide range of enteric lesions, mucosal ulcers were identified with a sensitivity of 81% for P1 lesions and 94% for P2 lesions, presenting impressive values of AUC (0.99 and 1.00, respectively). A recent study by Afonso et al.,⁴⁹ focused on risk potential assessment of small-bowel ulcers and erosions, achieving an accuracy rate of 95.6% in the detection and classification of erosions and ulcers, with any bleeding potential.

AI APPLICATIONS IN PCC

Two studies were conducted to evaluate deep learning performance among patients with CD undergoing PCC.^52,53 Majtner et al.⁵² used two splitting methods: random one and per-patient one (in which each patient’s frame was used only for training, validation, or testing). Only four of 558 images of the colon were misclassified as the small-bowel, and only seven of 1,000 images of the small-bowel were misclassified as the colon. The accuracy rates for the per-patient split and the random-split were 98.4% and 98.6%, either for small-bowel or colon lesions. Ferreira et al.,⁵³ demonstrated an impressive performance in the detection of ulcers and erosions in patients undergoing PCC (accuracy rate, 98.8%; negative predictive value, 99.5%), as well. Considering the average rate of 68 frames per second, it was estimated that only 12 minutes would require for a full PCC video revision.

FUTURE CHALLENGES USING THE AI FOR VCE

Though AI performance in patients undergoing VCE is impressive, there are still several challenges to be address in future studies. First, all but a single study were retrospective,⁵⁴ limiting the ability to explore performance in a real-life practice. Second, as some of the recent published studies focused on the classification of distinct parameters in CD lesions, it is of prime importance to further discriminate lesion features to improve prognosis predication, and accordingly personalize disease management. Third, as per-patient rather than per-lesion analysis better reflects a real-life practice, further adjustment and fine-tuning of CNN models are needed to cope with indistinguishable features, among frames from a single patient, to improve the accuracy rate of lesion classification. Fourth, comparative researches and accuracy-thresholds standardization should be addressed as well as demonstration of clinical correlation before it will be implemented in real-life practice.⁵⁵ Fifth, though capsule readers are generally eager toward AI-based VCE reading and interpretation, a substantial part of them are frightened of its implementation in real-life practice.⁵⁴ Moreover, almost half of capsule readers are aware of using AI application in medical fields.⁵⁴ Thus, a learning program training addressing it, is of prime importance among this population.^54,55 Sixth, assessment of its cost-effectiveness should be performed prior to AI implementation to real-life practice.⁵⁴ Finally, several technical adjustments might improve CNN model performance: (1) the current datasets composed of selected still frames, rather than video films,⁵⁵ with inherited risk of selection bias; (2) external validation datasets are mostly absent in the current studies,⁵⁴ limiting generalization of the studies’ findings; (3) stringent bowel-preparation protocols should be implemented to better cope with the low (but still exists) missed diagnosis rate using AI in VCE; (4) most of the studies have been dealt with a single type of VCE, limiting the use of the examined models in other VCE types.⁵⁵ Further development of a single and universal model for all VCE types, will probably prompt the incorporation of AI-based VCE reading in real-life clinical practice.

CONCLUSIONS

VCE has a crucial role in the management of patients with CD, to enable a reliable evaluation and monitoring of patients with active, as well as quiescent disease. The introduction of the colon capsule and subsequently the PCC allow a visualization of the entire small and large bowel during a single procedure, to facilitate disease management in those patients. The highest yield in determining disease extent and severity in a feasible and single procedure may improve patients’ adherence, mainly in patients with long-standing disease. Furthermore, the PCC has a prime benefit to precisely stage disease severity and extent, leading to therapy optimization and better clinical outcomes. The superiority of PCC in the detection of small-bowel lesions compared with MRE and in the same single procedure to identify colon involvement is promised in this field. As mentioned, severe adverse events are rare, mostly preventable by PC ingestion before VCE procedure.

Finally, the recent developments of machine learning applications in the detection and the grading of small and large bowel lesions (i.e., ulcer and its severity, erosion, stricture, and assessment of bleeding potential) have led to excellent performance and high accuracy rates as detailed above. Using the CNN models may shorten VCE reading time (up to 95%⁵⁵), resulting in a less tedious process with a potential to minimize missed diagnosis and false positive rates. However, the published literature in this field, a part of a single study was retrospective, which limits the ability to assess it in real-life practice, as well as the lacking cost-effectiveness evaluation. Also, to the best of our knowledge, no study has been conducted to assess AI-based VCE reading to predict future clinical outcomes in patients with CD. Considering the efficient and rapid process of AI-based VCE reading, developing of prediction-model using CNN architecture may substantially improve disease management of patients with CD, to afford treatment tailoring in this population.