Multimodal artificial intelligence in capsule endoscopy: Integrating video and sensor data for advanced gastrointestinal diagnostics

Abstract

Capsule endoscopy (CE) enables noninvasive visualization of the gastrointestinal tract but generates tens of thousands of images per examination, creating substantial interpretive burden, reader fatigue, and uncertainty in lesion localization. Artificial intelligence (AI), particularly deep learning-based image analysis, has markedly improved lesion detection and reduced reading time; however, image-only models remain limited in anatomical localization and spatiotemporal context. This narrative review summarizes current evidence on multimodal AI in CE, focusing on systems integrating capsule video with sensor-derived data such as inertial, magnetic, and physiological inputs. A structured search was conducted. Image-only convolutional neural network models consistently demonstrated high sensitivity (> 95%) for detecting gastrointestinal bleeding, ulcers, vascular lesions, polyps, and neoplasia, while reducing mean reading time from approximately 40-60 minutes to under 5 minutes. Multimodal architectures-including convolutional neural network-long short-term memory hybrids and sensor-fusion frameworks-further enabled accurate organ classification, transit time estimation, and improved lesion localization, particularly under suboptimal image quality. Overall, image-only AI excels in visual lesion detection, whereas multimodal AI enhances spatial awareness and diagnostic context. Key challenges include sensor heterogeneity, limited prospective validation, and regulatory considerations. With standardization and large-scale validation, multimodal AI may enable more context-aware and clinically interpretable CE.

Key Words: Capsule endoscopy; Artificial intelligence; Multimodal artificial intelligence; Deep learning; Convolutional neural networks; Gastrointestinal diagnostics; Data fusion; Lesion detection

Core Tip: Capsule endoscopy (CE) generates thousands of images per study, creating diagnostic and workflow challenges due to manual interpretation and localization errors. The integration of multimodal artificial intelligence combining visual data with sensor inputs such as inertial measurement units, magnetic trackers, and physiological monitors has significantly improved lesion detection, localization, and reading efficiency. Advanced architectures achieve sub-millimeter localization accuracy and > 95% diagnostic precision. These developments represent a paradigm shift in CE, transforming it from a passive imaging tool into an intelligent, context-aware diagnostic platform with the potential to enhance accuracy, reduce reading time, and standardize interpretation across clinicians.

INTRODUCTION

Capsule endoscopy (CE) has emerged as a noninvasive method for examining the gastrointestinal (GI) tract. It is used to diagnose various small bowel and large bowel diseases, including Crohn’s disease, celiac disease, suspected small bowel hemorrhage or lesions, and nonsteroidal anti-inflammatory drug-induced enteropathy in both the small and large bowel. It also enables visualization of areas in the jejunum and ileum that traditional endoscopy could not reach or access easily, thus expanding the diagnostic capabilities for GI disorders[1-4]. However, a single CE generates approximately 10000 images in one session, requiring over 40-50 minutes of reading time for each frame, proving a tedious task for gastroenterologists[5]. It is also highly dependent on the experience of the practicing gastroenterologist[6]. Therefore, this results in significant intraobserver and interobserver variability in the detection of vascular lesions, ulcers, and neoplasms as they could be limited to fewer frames, thereby increasing their risk of being overlooked[1,7].

To overcome these drawbacks artificial intelligence (AI) algorithms are being integrated with CE, thereby improving the detection of abnormalities and enhancing both sensitivity and specificity[8]. AI is defined as the ability of a computer to perform cognitive tasks comparable to those of a human. In recent years AI has been increasingly utilized in radiology, gastroenterology, orthopedics, and other medical fields to support healthcare professionals in diagnosing multiple conditions[9]. The deep learning algorithm is a form of AI that utilizes multiple neural networks arranged in layers to structure data in an orderly manner and extract information from it without requiring human assistance. In particular, convolutional neural networks (CNNs) are commonly used in deep learning models for image recognition and classification[10].

Despite these successes the drawback of analyzing image data alone is that it can be confounded by several artifacts, such as poor lighting in the GI tract, food debris, and uncontrolled peristaltic motion, thus hampering image quality[2,11]. More importantly, conventional AI cannot reliably determine the exact location of a lesion within the GI tract even though this information is crucial for clinical management[12]. Recent studies that incorporated multimodal AI in CE have been able to overcome these drawbacks by combining visual, sensory, and motion data to localize abnormalities in the GI tract with enhanced accuracy[13]. Figure 1 illustrates how multimodal AI combines capsule video with sensor data to improve lesion localization and diagnostic accuracy. To contextualize the evolution this review discusses the use of both traditional and multimodal AI in CE.

Figure 1 Components of multimodal artificial intelligence. This schematic illustrates how modern capsule endoscopy integrates multiple data streams to enable multimodal artificial intelligence (AI). As the capsule traverses the gastrointestinal tract, it captures video data and generates additional sensor-derived signals (e.g., localization, motility metrics, transit time, pH). These heterogeneous inputs are processed through multimodal AI frameworks that combine visual and non-visual modalities to enhance lesion detection, anatomical localization, motility assessment, and overall diagnostic accuracy. AI: Artificial intelligence. Created in BioRender.

BACKGROUND AND CURRENT STATE OF CE

CE was first introduced in 2001 and has since become an integral part of diagnosing diseases of the small intestine[14]. It involves patients swallowing a microcapsule, essentially a camera that captures images of the GI tract and transmits to a device located outside the patient’s body[14,15]. In clinical practice CE is widely used for small bowel evaluation because it is painless, requires no anesthesia, and allows visualization of otherwise obscure regions[9,15]. It is also used for screening and surveillance in familial adenomatous polyposis, Barrett’s esophagus, and esophageal varices and is increasingly explored for triaging and managing acute GI bleeding in the emergency and inpatient settings. In the colon it is reserved for incomplete colonoscopy or patients at high sedation risk and is contraindicated in cognitive impairment, high retention risk, and active Crohn’s disease with caution advised in those with cardiac implantable electronic devices[2].

Since the development of CE, several variations have entered the market, offering improved image resolution, wider field of view, improved battery life, and variable frame rates from 2 to 35 images captured by the capsule per second[16]. These include the PillCam SB3, Olympus Endocapsule 10, Micro Cam, and OMOM, all of which offer similar diagnostic accuracy[2,17]. The newer hybrid and technology-driven capsules, particularly the magnetically controlled gastric CE, enabled better evaluation by controlling capsule movement through the GI tract via magnetic steering[18]. However, they are expensive and complex to operate[16].

Although CE has several advantages, examining large volume images often results in lengthy reading time. Additionally, the accuracy in the detection of abnormalities is less because of the length of the small intestine and peristaltic motion, which can affect image quality. This not only strains physician time but also risks fatigue-related misses[9].

Therefore, in CE the integration of AI has enabled efficient analysis of vast volumes of visual data, assisting gastroenterologists in rapidly identifying clinically significant findings. Newer generations of high-definition capsules further enhance this capability by providing richer datasets for training CNNs, incorporating features such as the suspected blood indicator, adaptive frame rate technology, and the quick-view algorithm, which facilitates rapid detection of masses or bleeding sources[9,17,19]. In addition, modern capsule endoscopes now incorporate inertial and magnetic sensors that provide real-time positional and motion data, enabling more accurate localization beyond visual cues alone[20]. By integrating these multimodal data streams through AI, CE achieves improved diagnostic precision and spatial awareness, reflecting the evolution toward data-integrated medical AI systems[11,13].

ROLE OF AI IN CE

AI-based systems have begun to revolutionize medical imaging by performing initial interpretations and flagging abnormalities for clinicians, thus acting as a second pair of eyes[21]. CNNs are a specialized class of artificial neural networks that have become dominant in computer vision applications, including medical imaging[22]. Artificial neural networks form the foundational structure of both machine learning and deep learning and consist of multiple layers of interconnected “neurons” (mathematical functions) that learn complex patterns from data[23]. When trained on large, labeled datasets, such as CE image frames depicting ulcers, bleeding, or normal mucosa, the network iteratively adjusts its internal weights to achieve high classification accuracy[24,25]. Once trained, the model can categorize newly encountered images into clinically relevant groups (e.g., bleeding, ulcers, polyps), thereby reducing the cognitive burden associated with manual frame-by-frame review and improving overall sensitivity[26].

AI-based reading has also helped novice readers detect pathological lesions with significantly greater sensitivity than those without AI assistance and has brought their work performance and efficiency on par with those of experts. This performance advantage was demonstrated in studies by Ding et al[27] and Xie et al[28] in which junior readers supported by AI achieved significantly higher sensitivities for small bowel lesion detection (99.2% and 96.7%) compared with expert readers using conventional interpretation alone (91.1% and 88.8%, respectively). Notably, Ding et al[27] also reported a 33.3% reduction in missed diagnoses among novice readers when AI assistance was used, underscoring the role of AI in narrowing the expertise gap. This implies that AI can function as a built-in tutor or fail-safe, ensuring that less experienced physicians do not miss critical findings, thus standardizing CE interpretation quality across different centers and practitioners, an appealing prospect for healthcare systems aiming to deliver uniform care.

Additionally, AI can accelerate the reading process. A meta-analysis by Cortegoso Valdivia et al[26] demonstrated the superiority of AI-assisted reading over conventional interpretation in small bowel CE with significantly higher diagnostic accuracy and sensitivity. Across multiple studies AI-assisted reading achieved sensitivity values of 1.00, 0.99, 0.93, and 0.98 compared with 0.75, 0.88, 0.79, and 0.89 for conventional reading. AI assistance also markedly reduced the number of images requiring review, resulting in a dramatic reduction in mean reading time from 56.7 minutes to 4.7 minutes[26]. In a recent multicenter prospective study on small bowel bleeding, integrating an AI system into routine reading reduced the average reading time from 33.7 minutes to 3.8 minutes without compromising the diagnostic yield. This efficiency gain not only improves workflow but also potentially allows earlier decision-making[29].

MULTIMODAL AI IN CE: INTEGRATING VIDEO WITH SENSOR DATA

Multimodal fusion paradigms

Multimodal AI systems fundamentally differ from conventional image-only CNN models in CE analysis by fusing diverse data streams across different levels of integration. Integration can occur at different stages of the AI pipeline, generally categorized into three principal fusion strategies: Early; late; and hybrid. Figure 2 illustrates these fusion paradigms conceptually along with the strengths and weaknesses of each approach (strengths, weaknesses, opportunities, and threats analysis in Figure 3)[30].

Figure 2 Multimodal fusion paradigms in capsule endoscopy. Schematic representation of three primary strategies for integrating visual and sensor data in capsule endoscopy. A: Early fusion (e.g., Endo-VMFuseNet) combines visual and sensory inputs at the feature level using long short-term memory networks, achieving sub-millimeter localization accuracy without explicit calibration; B: Late fusion performs modality-specific analyses independently, then merges model outputs via weighted voting or averaging to produce a robust final prediction; C: Hybrid fusion (e.g., convolutional neural network-long short-term memory hybrid) integrates spatial features from video with temporal features from inertial measurement unit data, enabling accurate organ localization and transit-time estimation (> 95% accuracy). LSTM: Long short-term memory; IMU: Inertial measurement units. Created in BioRender.

Figure 3 Strengths, weaknesses, opportunities, and threats analysis of multimodal artificial intelligence systems for capsule endoscopy. This figure illustrates the strengths, weaknesses, opportunities, and threats associated with integrating multimodal artificial intelligence into capsule endoscopy. Key advantages include enhanced lesion localization, diagnostic accuracy, and workflow efficiency. Challenges span computational demands and limited clinical validation. The approach offers opportunities for advanced lesion mapping and integration of diverse sensor data, while regulatory, interoperability, and data security concerns represent potential barriers. GI: Gastrointestinal; CE: Capsule endoscopy; IBD: Inflammatory bowel disease.

Early fusion (feature-level integration): In early fusion raw or extracted features from video and sensor data are concatenated prior to being input into a unified model. This approach allows simultaneous learning from visual and non-visual cues. A representative model is Endo-VMFuseNet, which employs CNNs to process visual odometry and magnetic sensor inputs, subsequently merging them through hierarchical long short-term memory (LSTM) networks. This architecture achieved sub-millimeter localization accuracy without requiring explicit sensor calibration or synchronization, thereby addressing one of the major challenges in capsule localization[31].

Late fusion (decision-level integration): Late fusion involves analyzing each modality independently and subsequently combining its outputs via weighted voting or averaging to generate final predictions. This strategy enhances system robustness because performance degradation in one modality, such as temporary sensor dropout, does not drastically impair the final decision, improving reliability in real-world applications[32].

Hybrid fusion (multistage integration): Hybrid fusion combines the strengths of both early and late fusion. Here, modality-specific features are first extracted individually and then integrated at a later stage for contextual refinement. Based on this, Nam et al[11] developed a CNN–LSTM hybrid network that fuses spatial features from CE video with temporal inertial measurement unit data. Their system demonstrated over 95% accuracy in classifying organ locations and estimating GI transit times.

In practice the choice of fusion strategy depends on the application. Early fusion suits tightly linked data in which early interactions matter. Late fusion is useful when one data source may be unreliable, and hybrid fusion is increasingly favored as a balanced approach that combines both. Across all methods, accurate time alignment of data streams and robustness to noise remain key challenges[33,34].

CLINICAL APPLICATIONS

The clinical utility of AI-driven CE extends across a broad spectrum of GI pathologies in which its ability to enhance diagnostic accuracy, efficiency, and reproducibility has been repeatedly demonstrated. Multiple studies have evaluated AI-driven systems in the detection of GI hemorrhage, erosions, ulcers, vascular lesions, polyps, and tumors, conditions that often pose interpretative challenges during manual CE review.

GI hemorrhage

GI bleeding is the most common anomaly of the GI tract and a cardinal manifestation in conditions such as ulcers, polyps, tumors, and inflammatory bowel disease. This makes the detection of GI bleed an essential component of CE[27,35]. AI systems have demonstrated excellent pooled diagnostic performance for GI ulcers and hemorrhage with an overall accuracy of 95.4%, sensitivity of 95.5%, and specificity of 95.8% and even higher performance for hemorrhage detection alone (sensitivity 96% and specificity 97%)[36,37]. Jia et al[38] demonstrated superior precision and recall with a CNN model compared with conventional bleeding detection methods, along with a higher F1 score (a composite metric defined as the harmonic mean of precision and recall that reflects the optimal balance between false-positive and false-negative classifications). A higher F1 score indicates more reliable and clinically robust bleed detection, particularly in class-imbalanced settings such as CE in which accuracy alone can be misleading[38,39]. Similarly, in a multicenter prospective study by Spada et al[35] integrating AI for suspected small bowel bleeding significantly increased diagnostic yield at the patient level while slashing the mean reading time from 33.7 minutes to approximately 4 min (a nearly 10-fold reduction).

Erosion and ulcers

Erosions and ulceration are the most commonly encountered abnormalities in CE[25,35]. They can be subtle and numerous, making them tedious to identify manually. AI algorithms have excelled here as well. For example, Ribeiro et al[40] designed and tested a CNN model for detecting colonic ulcers and erosions in colon CE images, achieving a sensitivity of 96.9% and a specificity of 99.9%. While that study was in the colon, similar approaches in small bowel CE have also shown high sensitivity for ulcer detection. Aoki et al[41] presented a deep learning-based CNN model that demonstrated high accuracy for automatic detection of small bowel erosions and ulcerations on CE, achieving an area under the curve of 0.958 with 88.2% sensitivity and 90.9% specificity.

Vascular lesions

Angioectasias (vascular malformations) and varices appear as faint red spots and can be easily overlooked. Mascarenhas et al[42] conducted a multicentric study for the detection of vascular lesions based on a set of 1022 CE procedures. It was the first proof-of-concept CNN model for pan endoscopic detection of vascular lesions. The results showed 95% overall diagnostic accuracy with a sensitivity and specificity of 86.4% and 98.3%, respectively. Similarly, a CNN-based model for CE demonstrated by Ribeiro et al[43] showed an excellent performance for automatic detection of small bowel vascular lesions, achieving a sensitivity of 91.8%, specificity of 95.9%, and overall accuracy of 94.4% for identifying red spots, angioectasias, and varices. Notably, the algorithm processed images at a rate of 145 frames per second, enabling near–real-time analysis and markedly improving reading efficiency while maintaining high diagnostic precision[43]. This level of accuracy is promising because vascular lesions often cause occult bleeding. Therefore, an AI that consistently flags them could improve diagnostic yield for obscure bleeding cases.

Polyps and tumors

CE is also used for colorectal cancer screening in patients who cannot undergo colonoscopies (e.g., due to anesthesia risk). The second-generation colon CE-2 has reported a sensitivity of 85%-87% and a specificity of 85%-95% for the detection of colorectal polyps ranging in size from < 6 mm to ≥ 10 mm[44]. A systematic review by Moen et al[45] reported sensitivities ranging from 47.4% to 98.1% and specificities of 87.0% to 96.3% in per-frame analysis using proposed AI designs for polyp or colorectal neoplasia detection[46-48]. The two studies that performed per-lesion analysis showed significantly improved sensitivity of polyp or colorectal neoplasia detection to 81.3%-98.1%[45,46]. All these are summarised in Table 1.

Table 1 Clinical applications of artificial intelligence in capsule endoscopy¹.

Lesion type	Ref.	Model	Capsule type	Number of training images	Performance metrics
GI hemorrhage	Jia et al[38]	Deep CNN for bleeding detection	Small bowel	10000 images (bleeding and non-bleeding)	Improved precision, recall
	Spada et al[35]	A multicenter study utilizing AI-assisted reading of lesions	Small bowel	158235 images	Reading time reduced: 33.7 min → 3.8 minutes (P < 0.0001); improved accuracy
Erosions and ulcers	Ribeiro et al[40]	CNN model for colonic ulcers	Colon capsule	37319 (3570 with lesions)	Sensitivity: 96.9%, specificity: 99.9%
Vascular lesions	Mascarenhas et al[42]	Multicenter study utilizing CNN	Small bowel and colon capsule	34665 (11091 with lesions)	Diagnostic accuracy: 95%, sensitivity: 86.4%, specificity: 98.3%
Polyps and tumors	Kjølhede et al[44]	Systematic review and meta-analysis for the detection of polyps < 6 mm to ≥ 10 mm	CCE-2	Combined across studies	Sensitivity: 85%-87%, specificity: 85%-95%
Polyps and tumors	Moen et al[45]	Systematic review of AI models for polyp or colorectal neoplasia detection	CCE-2	Varied across studies (thousands to 30000)	Per-frame sensitivity: 47.4%-98.1%, specificity: 87.0%-96.3%; per-lesion sensitivity: 81.3%–98.1%

¹Summary of key studies evaluating the diagnostic performance of artificial intelligence (AI)-assisted capsule endoscopy across various gastrointestinal lesion types. Convolutional neural network-based models have demonstrated improved precision, sensitivity, and specificity in detecting gastrointestinal bleeding, erosions, ulcers, vascular lesions, and colorectal polyps compared with conventional reading. AI integration has also significantly reduced reading time and enhanced diagnostic consistency among readers of different experience levels.

CCE: Colon capsule endoscopy; CNN: Convolutional neural networks; AI: Artificial intelligence.

COMPARATIVE PERFORMANCE OF IMAGE-ONLY AND MULTIMODAL AI SYSTEMS

Across diagnostic contexts, image-only CNNs consistently achieve excellent lesion-level sensitivity (95%-99%) for bleeding and ulcer detection, but their localization accuracy remains limited. In contrast, multimodal systems, such as CNN-LSTM and EndoVMFuseNet, demonstrate marked gains in spatiotemporal mapping and organ classification (> 95% accuracy, sub-millimeter localization) even when visual quality is suboptimal[11,49]. These results suggest that image-only AI excels in feature-intensive, high contrast lesions (e.g., active bleeding), whereas multimodal AI confers the greatest advantage in complex tasks requiring anatomical context or motion tracking, such as transit time estimation or three-dimensional trajectory reconstruction[49]. Thus, performance superiority depends on the diagnostic task rather than the architecture alone. A summary of the representative studies evaluating image-only and multimodal AI systems is presented in Table 2.

Table 2 Comparative overview of image-only vs multimodal artificial intelligence architectures in capsule endoscopy¹.

Model	Modality type	Capsule platform	Key diagnostic application	Performance accuracy
Ding et al[27], 2023 (CNN)	Image only	Small bowel capsule	Lesion detection (ulcers, bleeding)	Sensitivity 99.2%, specificity 96.7%
Nam et al[11], 2024 (CNN-LSTM)	Multimodal (video + IMU)	PillCam™ SB3	Organ localization and transit time estimation	> 95% accuracy
Turan et al[31], 2017 (Endo-VMFuseNet)	Multimodal (video + magnetic)	Experimental platform	Capsule localization (3D trajectory mapping)	Sub-millimeter accuracy
Mascarenhas et al[42], 2024 (CNN)	Image only	Multibrand CE	Vascular lesion detection	Accuracy 95%, sensitivity 86.4%, specificity 98.3%
Vedaei and Wahid[13], 2021 (prototype)	Multimodal (video + IMU)	Research prototype	3D trajectory reconstruction	Improved localization accuracy

¹Summary of representative studies evaluating artificial intelligence (AI) systems in capsule endoscopy, comparing image-only convolutional neural network (CNN) models with multimodal architectures that integrate additional sensor inputs such as inertial measurement units or magnetic trackers. Image-only CNN models show excellent lesion detection accuracy (95%-99%) for bleeding and ulcer identification, whereas multimodal systems, such as CNN-long short-term memory and Endo-VMFuseNet, achieve superior spatial localization and temporal mapping capabilities (> 95% accuracy, sub-millimeter precision). These findings highlight that while image-only AI excels in visual lesion recognition multimodal approaches provide enhanced anatomical context and diagnostic robustness.

LSTM: Long short-term memory network; 3D: Three-dimensional; CNN: Convolutional neural networks; IMU: Inertial measurement units; CE: Capsule endoscopy.

BENEFITS, IMPACT, CHALLENGES, AND FUTURE DIRECTIONS

Multimodal AI integration significantly enhances CE diagnostics by providing precise lesion localization, improved detection sensitivity, and dramatically reduced physician review time[11,50,51]. However, practical implementation faces challenges, including sensor misalignment, drift, computational complexity, and limited large-scale clinical validation[11,13]. Future research directions include adopting advanced AI architectures such as transformer models and graph neural networks, optimizing lightweight on-device models, and conducting comprehensive multicenter trials to validate clinical efficacy across diverse patient populations and CE platforms[13,50].

Beyond current architecture, transformer-based multimodal fusion (e.g., Vision Transformers and Swin-Transformer hybrids) and self-supervised pretraining approaches may enable models to learn cross-modal representations from fewer labeled datasets[52]. The integration of edge AI hardware could further enable real-time, low-latency inference for lesion detection and localization. Future work should also explore adaptive AI pipelines capable of online learning from clinician feedback, advancing CE from static analysis toward autonomous, context-aware diagnostics.

CHALLENGES AND LIMITATIONS

While multimodal AI in CE offers transformative potential, several challenges and limitations hinder its seamless clinical adoption. One major challenge is the generalizability of AI models. Many current algorithms are trained on retrospective, often single-center datasets that may not capture the full diversity of real-world patient populations or endoscopic findings. This can result in spectrum bias and reduced performance in broader clinical practice[53].

Technical limitations are also significant. CE images are frequently affected by motion artifacts, suboptimal bowel preparation, and variable lighting, all of which can impair AI accuracy and reliability. Sensor-based localization while vital for mapping lesions remains susceptible to signal loss, interference, and a lack of standardization across different platforms[54]. Furthermore, most AI systems have not been validated in large, prospective, multicenter trials, raising concerns about overfitting and investigator bias[55].

Ethical and legal considerations are increasingly important as AI systems become more autonomous. Issues such as patient data privacy, algorithmic bias, and the allocation of responsibility for diagnostic errors remain unresolved[56]. The economic burden of implementing advanced AI and capsule technology, including costs for devices, data storage, and computational infrastructure, may also restrict access, especially in resource-limited settings.

Finally, despite advances in automation expert human oversight remains essential. False positives and negatives can occur, and AI should currently be viewed as an adjunct rather than a replacement for clinical expertise[57].

CONCLUSION

Multimodal AI represents a significant advancement in CE by extending analysis beyond visual pattern recognition to incorporate spatial, temporal, and contextual information derived from auxiliary sensor data. Image-only deep learning models have consistently demonstrated excellent performance in lesion detection, particularly for bleeding, ulcers, vascular abnormalities, and neoplasia, while substantially reducing physician reading time. However, their reliance on visual cues alone limits anatomical localization and interpretability in dynamic GI environments.

By contrast, multimodal architectures that integrate video with inertial, magnetic, or physiological sensor inputs provide enhanced lesion localization, organ classification, and transit time estimation, addressing long-standing limitations of conventional CE. These capabilities support more precise clinical decision-making, facilitate targeted downstream interventions, and improve diagnostic confidence, particularly in complex or equivocal cases. Importantly, the complementary strengths of image-only and multimodal systems suggest that performance gains are task-dependent rather than architecture-dependent.

Despite these advances, widespread clinical adoption remains constrained by challenges related to data heterogeneity, platform-specific sensor variability, limited prospective multicenter validation, and unresolved ethical and regulatory considerations. Continued emphasis on standardized benchmarking, transparent model development, and large-scale clinical trials will be essential to ensure safe, reproducible, and equitable implementation.

With ongoing technological refinement and appropriate clinical validation, multimodal AI has the potential to transform CE from a passive imaging modality into an intelligent, context-aware diagnostic platform that enhances accuracy, efficiency, and consistency of GI disease evaluation.