List of Papers Browse by Subject Areas Author List
Abstract
Video capsule endoscopy (VCE) is the preferred first-line modality for visualizing the small bowel mucosa and evaluating obscure gastrointestinal bleeding. However, each examination generates over 50 thousand frames, making manual review time-consuming and prone to missed findings and inter-observer variability. Most AI systems rely on single-frame analysis, overlooking temporal dynamics and limiting their ability to distinguish anatomical transitions from transient artifacts. We propose a real-time auditing framework that jointly identifies the anatomical region along the gastrointestinal tract and detects anomalies through intrinsic temporal modeling. Consecutive frames are encoded using an endoscopy-pretrained CNN and processed by a lightweight Vision Transformer operating in a sequence-to-sequence manner, enabling frame-level predictions without post-processing. Evaluated on the multi-center Galar dataset (~3.5 million frames), the proposed system achieves a macro F1-score of 90% for anatomical classification and 67% macro recall for binary anomaly detection within a unified multi-task learning framework, outperforming prior frame-wise approaches. Deployment on a low-power edge device confirms real-time performance (30.5 ms per frame, ~33 FPS), enabling continuous monitoring of incomplete procedures, capsule retention, and persistent abnormalities. Our code is available at https://github.com/DiegoBravoH/CapsuleAudit.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0867_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/DiegoBravoH/CapsuleAudit
Link to the Dataset(s)
Galar: [https://plus.figshare.com/articles/dataset/Galar_-a_large_multi-label_video_capsule_endoscopy_dataset/25304616/1](https://plus.figshare.com/articles/dataset/Galar-_a_large_multi-label_video_capsule_endoscopy_dataset/25304616/1)
Rhode Island Gastroenterology Video Capsule Endoscopy Dataset: https://springernature.figshare.com/collections/Rhode_island_gastroenterology_video_capsule_endoscopy_data_set/6071216/1
BibTex
@InProceedings{BraDie_ComputerAssisted_MICCAI2026,
author = { Bravo, Diego AND Ruano-Balseca, Josué AND González, Fabio A. AND Romero, Eduardo},
title = { { Computer-Assisted Intervention in Capsule Endoscopy: A Real-Time Edge-AI Auditing System } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a real-time auditing framework for Video Capsule Endoscopy (VCE) to reduce the manual burden of reviewing 50,000+ frames per patient. The authors introduce a sequence-to-sequence temporal modeling approach to detect both anatomical structures and anomaly. The system uses a pretrained ResNet-50 for feature extraction followed by a lightweight Vision Transformer (ViT-Small) that treats frame embeddings as temporal tokens. This multi-task framework jointly predicts anatomical regions (5 classes) and binary anomalies. The model is deployed on a low-power NVIDIA Jetson Nano, achieving real-time performance (~33 FPS), which allows for intra-procedural monitoring of capsule retention and examination completeness.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1- Clinical and Practical Utility: Video Capsule Endoscopy is the primary non-invasive modality for visualizing the entire small bowel, and the authors tackle it with a temporal model designed specifically for efficiency. By focusing on edge deployment, they move beyond simple analysis and provide a framework that can be used in real-time to monitor the procedure.
2- Temporal and Multi-Task model: The pipeline is well-engineered, featuring a comprehensive ablation study on temporal window sizes (ranging from 5 to 195 frames) and a comparison between single-task and multi-task learning. The small size of both spatial and temporal decoders allow for temporal modelling with a high inference speed.
3- Hardware validation: The deployment on a Jetson Nano provides empirical evidence of the system’s efficiency. Providing concrete latency numbers on constrained hardware supports the “Edge-AI” claim.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1- Incremental Methodological Novelty: From a deep-learning perspective, the architecture is relatively standard. The novelty of the work lies primarily in its application and hardware deployment insights rather than in a fundamental methodological innovation.
2- Lack of deployment ablation: While the edge deployment is one of the strengths of the paper, it would have been nice to have an ablation of the single frame models vs temporal models with different temporal windows. This would highlight the efficiency-performance trade-off.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper provides a practical and well-executed solution for real-time video capsule endoscope auditing. While the methodological novelty is limited, the work is valuable due to its focus on edge deployment and clinical utility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
As an application paper focusing on capsule endoscopy, a highly relevant clinical problem, it provides a well-executed, end-to-end pipeline, ranging from training spatiotemporal models with competitive performance to deployment on a real-time edge device.
Review #2
- Please describe the contribution of the paper
The paper proposes a spatio-temporal deep learning method for simultaneous recognition of anatomy and anomalies in video capsule endoscopy. Deployed on a low-power AI-chip, it is shown that the method can process the clinical videos in real time.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper poses a clinically relevant topic with the potential to reduce the time clinicians need to spend reviewing the endoscopic videos. 2.Ablations and comparisons to related work are performed on a public dataset, showing increased performance for anomaly detection and good performance for anatomy detection. The latter task is validated on an additional dataset, without fine-tuning, to better reflect real clinical performance. 3.Showcasing that the method can perform detection in real-time, by implementation on a low-power AI-chip, is a good effort in closing the bridge to clinical implementation.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.It is difficult to assess the effects of window size and STL/MTL on performance in anatomy classification, due to the dense presentation in Table 1.Therefore, the selection of MTL 75 as the best-performing is not obvious. Providing more (visual) guidance would help the reader follow the paper’s intermediate conclusions. Additionally, the text under Table 2 reports a recall/precision/F1 of 77%/81%/79%, which is not consistent with the numbers presented in the table. 2.Although the proposed method is sound and follows standards in the field, it does not bring unique components designed particularly for the task at hand. 3.If I understood correctly, the Galar dataset provides more labels (such as good/reduced views), which I would have expected to be part of a multi-task method tailored for the task at hand. 4.While the clinical use of anomaly detection is clear, it would be good to explain more about the clinical use of automated anatomy classification. Without it, it is more difficult to assess the relevance of its results.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a method for a clinically relevant topic, showing clear improvements over related work. Although the methodology and experiments are solid, the presentation of results is sometimes unclear (see comment 1). After the necessary clarifications and adaptations, I think the paper could potentially be accepted.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors provided satisfactory answers to my raised concerns. Overall, the work presents a valuable solution to a real clinical challenge, validated on edge hardware. As the authors state in their rebuttal, the explanation for selecting MTL-75 as the best-performing model, together with the metric presentation in Section 3.2, will be improved in the final manuscript.
Review #3
- Please describe the contribution of the paper
This paper presents a real-time edge-AI system for video capsule endoscopy (VCE) that jointly performs anatomical localization and binary anomaly detection.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Addresses a clinically important and practical VCE workflow problem. 2.Includes real-time deployment on a low-power edge device, which is a clear practical strength. 3.Moving beyond frame-wise prediction and HMM post-processing is a meaningful direction.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The sequence-to-sequence temporal modeling is under-described. This is the core contribution, but the paper gives too little detail on how the temporal modeling actually works in practice. It remains unclear how temporal order is encoded, how causal inference is implemented, and what specifically makes the approach more than a straightforward transformer over frame embeddings. 2.The baseline selection is too limited for a paper centered on temporal modeling. The experiments mainly compare against medical/VCE-specific frame-wise methods. I recommend including representative video anomaly detection & classification / general temporal modeling baselines (also higher-level conferences and journals), rather than restricting comparisons to medical methods only.[1-3] 3.The effect of foundation-model pretraining is not sufficiently disentangled from the proposed method. From Table 1, the pretrained weights seem to contribute substantially to the final gains. In this case, comparing only to ImageNet-pretrained models is not enough (too old). The paper should include stronger modern natural-scene foundation models, e.g. DINOv2, to better isolate the value of the temporal design from the value of initialization. 4. [1] Wu, Peng, et al. “Deep learning for video anomaly detection: A review.” IEEE Transactions on Neural Networks and Learning Systems (2026). [2] Ye, Muchao, Weiyang Liu, and Pan He. “Vera: Explainable video anomaly detection via verbalized learning of vision-language models.” Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.[3] Zhou, Hang, et al. “Video anomaly detection with motion and appearance guided patch diffusion model.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39.No. 10.2025.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is based on the mismatch between the paper’s practical importance and the current level of technical and experimental support. The application is clinically relevant and the edge deployment results are a clear strength. However, the main contribution, sequence-to-sequence temporal modeling, is not described in enough detail to be fully convincing. The experimental baselines are also too limited for a paper centered on temporal modeling, especially without comparisons to broader video/temporal methods. In addition, a substantial part of the gain may come from pretrained initialization, but this is not well disentangled, and comparison only to ImageNet-pretrained models is insufficient. Overall, the paper is promising, but the current evidence is not yet strong enough for acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author has answered some of my concerns, and I think it is worth accepting now.
Author Feedback
The authors thank the reviewers for their insightful comments. [R2, R3, R4] Novelty, Temporal Modeling, and Contribution The contribution of this work is primarily a computer-assisted intervention system: the novelty lies in enabling real-time temporal VCE analysis under strict edge-computing constraints through a lightweight framework, a setting insufficiently explored by existing approaches. Under these constraints, the architecture integrates established backbones for causal online inference: (1) sliding windows contain only past and present frames, consistent with real-time deployment; each frame is encoded once by ResNet50 and reused across overlapping windows (stride = 1) through a ViT-Small temporal encoder for per-frame prediction; (2) temporal order is encoded via learnable positional embeddings, while frame predictions are aggregated through soft voting with entropy-based uncertainty; (3) homoscedastic MTL jointly optimizes anatomical localization and anomaly detection in a single edge-deployable model (30.5 ms/frame on NVIDIA Jetson Nano); and (4) the system is validated on a recent 2025 multi-center, multi-device dataset (~3.5M frames), including evaluation on an unseen dataset. We will clarify this framing in Section 1. [R4] SOTA Baselines and Pretrained Non-medical temporal baselines were not included because the primary scope of this study was real-time auditing of complete VCE procedures under edge-deployment constraints. A reviewer-suggested VLM contains >8B parameters, whereas our approach contains ~45M and runs fully online. VCE presents unique challenges, as the capsule moves through a continuously deforming organ without controlled navigation, while most temporal foundation models are trained in fixed-camera or operator-controlled settings. Benchmarking and fine-tuning such models for real-time VCE remains important future work. Selected temporal baselines previously applied to endoscopy video analysis included CNN+GRU and CNN+ViT models evaluated on VCE videos under the same protocol (Section 3.2). Regarding foundation-model pretraining, both CNN and ViT backbones were pretrained using self-supervised DINO on the GastroNet [15] dataset (>5M gastrointestinal images from 8 clinical centers). The proposed approach integrates these backbones for real-time VCE analysis under edge-deployment constraints, and results indicate improved performance (Table 1). The available GI-pretrained DINOv2 ViT-Base backbone (~86M parameters) [21] is not suitable for real-time deployment. [R3, R2] Results Clarity and Deployment Trade-offs The distinction between macro metrics (Table 2) and weighted metrics (text) reflects their complementary roles: macro metrics reflect performance across classes under severe imbalance, while weighted metrics account for class support. MTL-75 was selected to prioritize macro recall in the anomaly task. This configuration achieves 66.7% macro recall (Table 2), 77% weighted recall for the anomaly task (text), and 90.3% macro F1 for the anatomy task (Table 1) using a single model. We will better justify this selection in Subsection 3.2.Additionally, a systematic latency analysis across temporal window sizes and backbone architectures on edge devices is planned as future work. [R3] Clinical Relevance The anatomical classification task enables: (1) GI transit time and motility assessment; (2) detection of retrograde events associated with Barrett’s metaplasia [22]; and (3) early detection of capsule retention (0.7-1.4% of examinations). We will clarify the clinical utility in the final paragraph of the Results section. Additional Galar annotations (e.g., good/reduced-view) are only available for 6/80 studies. Since anatomical continuity is preserved across reduced-view frames, our temporal context supports inference under visual ambiguity, while entropy-based uncertainty identifies low-confidence predictions without explicit view-quality labels, a key advantage over frame-wise approaches.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The manuscript was reviewed by three expert reviewers. While acknowledging merits in the proposed work, the reviewers also raised some concerns about methodological novelty and validation. Such concerns should be addressed to strengthen the value of the work.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors provided satisfactory answers to the concerns raised by the reviewers.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents an application study for an automatic system for video analysis during capsule endoscopy. While the proposed approach does not present a strong methodological novelty, reviewers have highlighted the value of the implementation on an edge device and practical utility. I agree with the overall positive rating for acceptance by the reviewers.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Authors addressed reviewers’ concerns. I would recommend to include and discuss it in camera-ready.
