Pith. sign in

REVIEW 4 major objections 5 minor 79 references

BleedOrigin: Dynamic Bleeding Source Localization in Endoscopic Submucosal Dissection via Dual-Stage Detection and Tracking

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A dual-stage AI pipeline pinpoints and tracks ESD bleeding sources, from onset to hemostasis planning.

desk verdict The tracking work is plausible and the dataset is a real resource, but the onset-detection results are invalid as presented: the text says every test clip bleeds at frame 120, so a constant predictor would beat the reported numbers. read the letter →

arxiv 2507.15094 v1 pith:YFCRD3DG submitted 2025-07-20 cs.CV cs.AI

Mengya Xu , Rulin Zhou , An Wang , Chaoyang Lyu , Zhen Li , Ning Zhong , Hongliang Ren This is my paper · ORCID
classification cs.CVcs.AI
keywords endoscopicsubmucosaldissectionbleedingsourcelocalizationonsetdetectionpointtrackingsurgicalvideoanalysispseudo-labellearningparameter-efficientfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to make bleeding-source localization in endoscopic submucosal dissection (ESD) a solvable computer-vision problem rather than a clinician's visual hunt. It introduces BleedOrigin-Bench, a dataset of 106,222 frames from 44 ESD procedures with 1,771 expert-annotated bleeding sources, and BleedOrigin-Net, a two-stage framework that first detects the frame and pixel of bleeding onset and then tracks that point continuously. The reported results are 96.85% frame-level accuracy (within ±8 frames) for onset detection, 70.24% pixel-level accuracy (within 100 px) for initial source detection, and 96.11% pixel-level accuracy (within 100 px) for continuous tracking. If these numbers hold in deployment, an endoscopist could receive a real-time alert and a continuously updated pointer to the bleeding source, cutting the repeated water-flushing cycles that currently slow ESD procedures.

What carries the argument

The load-bearing machinery is the two-stage architecture plus the training signals that make it work in a data-sparse domain. BleedOrigin-Detect uses a Multi-Domain Confidence-based Frame Memory (MDCFM) module, which keeps only 'clean view' frames by thresholding RGB, HSV, and optical-flow differences, and Multi-Domain Gated Attention (MDG) to fuse those domains when comparing frames; a sliding-window transformer then predicts the onset time, and a multi-scale attention head over heatmap, red-mask, and perceptual-encoder features predicts the source coordinate. BleedOrigin-Track builds dense supervision for a transformer point tracker by matching high-confidence feature points near the annotated source, propagating them with the tracker, smoothing trajectories with a Kalman filter, and fine-tuning only low-rank adapters to avoid unstable full fine-tuning. A periodic memory-refresh strategy re-grounds the tracker every N frames to stop long-clip drift.

What would settle it

Run the full pipeline end-to-end on the three densely annotated deployment clips and measure the fraction of tracked frames within 100 px of the expert label; if that fraction falls substantially below the 96.11% reported for tracking with ground-truth initialization, the compounding assumption behind the clinical claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that the complete bleeding-management workflow in ESD can be covered by one dual-stage pipeline: a detection stage that pinpoints the first bleeding frame and the source coordinate, and a tracking stage that keeps that point under continuous observation through water flushing, camera jitter, light reflection, and instrument interference. The paper argues that this division of labor is what previous work misses, since bleeding-region segmentation and generic point trackers either ignore the source or cannot survive the ESD visual environment. On the new benchmark, BleedOrigin-Net outperforms YOLO-family detectors, multimodal large language models, and state-of-the-art point trackers on all three sub-tasks, with the largest margin in continuous tracking.

Load-bearing premise

The clinical claim of continuous guidance rests on the assumption that the three stages still perform at their reported accuracy when the predicted onset frame and predicted source coordinate, rather than the ground-truth ones, feed the tracking stage, which the paper tests only qualitatively.

Editorial extensions

If this is right

  • An ESD guidance system could alert the surgeon at bleeding onset and hold a pointer on the source through most of the episode, giving hemostatic tools a target even when blood pools obscure the view.
  • With 96.85% onset accuracy within ±8 frames and 96.11% tracking accuracy within 100 px, the remaining bottleneck is the initial source localization, which reaches only 70.24% within 100 px.
  • The tracking stage runs at 11.82 FPS on a standard GPU and the full pipeline at roughly 6 to 10 FPS per stage, so the approach is compatible with near-real-time use in the operating room.
  • The dataset's 8 anatomical sites, 6 challenging scenarios, and patient-level splits give the field a common benchmark for comparing future ESD bleeding-source methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported accuracies are stage-wise, with ground-truth initialization; the likely end-to-end accuracy on the deployment clips will be lower because onset-frame and source-coordinate errors from earlier stages seed the tracker, and a direct end-to-end measurement on the three densely annotated clips would quantify that drop.
  • The paper states both 1 FPS downsampling and 30 FPS frame sampling in different sections, so the clinical latency implied by the ±8-frame onset tolerance depends on which rate governs the detection clips; the 'real-time' claim should be read against the benchmark's actual temporal resolution.
  • The pseudo-label recipe (sparse endpoint matching, trajectory propagation, Kalman smoothing) is transferable to other surgical point-tracking problems with sparse annotations, such as following a tool tip or a vessel across a procedure.
  • A comparison against a simple segmentation-centroid baseline, which the paper does not report, would clarify whether the attention-based source detector adds value beyond locating the middle of the detected bleeding region.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces BleedOrigin-Bench, a new dataset of 485 bleeding clips from 44 ESD procedures with 1,771 expert-annotated bleeding-source frames and 39,755 pseudo-labeled frames, and BleedOrigin-Net, a two-stage pipeline: BleedOrigin-Detect predicts the bleeding onset frame and the initial source coordinate, and BleedOrigin-Track propagates that point over time using pseudo-label training plus LoRA/AdaLoRA fine-tuning of Track-On. The authors report state-of-the-art numbers: 96.85% frame-level accuracy (±8 frames) for onset detection, 70.24% pixel-level accuracy (≤100 px) for initial source detection, and 96.11% pixel-level accuracy (≤100 px) for continuous tracking. The paper also includes ablations, a comparison with YOLO, MLLM, and point-tracking baselines, and a qualitative deployment study on three full-length clips.

Significance. If the reported results hold, the dataset and pipeline would be a useful contribution to an underexplored clinical task, and the patient-level split, the explicit separation of evaluation and deployment modes, and the pseudo-label generation strategy with Kalman smoothing are sensible design choices. However, the onset-detection evaluation protocol described in Section 3.3.1 makes the headline detection result degenerate as reported, and the end-to-end clinical claim is only qualitatively supported. The dataset statistics also contain several internal inconsistencies. With corrections, the dataset itself could still be a valuable resource, and the tracking component shows a clear improvement over generic point trackers, but the current manuscript does not support the stated state-of-the-art claims without revision.

major comments (4)
  1. [Section 3.3.1 and Eq. (9)] The onset-detection benchmark is degenerate as described. The text states that every clip consists of 150 consecutive frames and that 'the initial bleeding source consistently occurs at frame 120,' with validation and testing using full clips without frame skipping. Under Eq. (9), a constant baseline that always predicts t_pred = 120 achieves Acc±k = 100% for every k ≥ 0 on the 98 test clips. The reported BleedOrigin-Detect-Frame results (Table 3: 49.92% at ±0 and 96.85% at ±8) are below this trivial baseline, so the claimed state-of-the-art performance for the first stage does not demonstrate detection capability as currently evaluated. Please clarify whether the test onset frames actually vary, or evaluate on full-length videos with naturally varying onset times; in either case, report a constant-frame baseline and the distribution of onset frames.
  2. [Section 4.3 and Algorithm 1] The quantitative evaluation is stage-wise and uses oracle initialization. The text explicitly states that the ground-truth initial bleeding frame is used to evaluate the initial source, and that the ground-truth initial bleeding frame and coordinate are used as the starting point for tracking evaluation. The end-to-end deployment mode is only assessed qualitatively on three clips (Section 6.2 and Section 6.4). Because the clinical claim is continuous guidance in real use, and because the onset detector is the first stage of that pipeline, the paper needs either a quantitative end-to-end evaluation on held-out clips with predicted initialization or a clearly stated limitation that only component-wise accuracy is established.
  3. [Sections 3.1, 3.2, 3.3.1, 3.3.2] The dataset description contains several inconsistencies that affect reproducibility. First, 44 patients are reported, but the train/validation/test split sums to 25 + 6 + 8 = 39 patients, and a 4:1:1 patient-level split of 44 patients would not produce those counts. Second, Section 3.1 says all videos are uniformly downsampled to 1 FPS, while Section 3.2 says frames are systematically sampled at 30 fps to obtain 106,222 images. Third, Section 3.3.2 says Long Clips range from 31 to 731 frames, while Section 3.3.1 defines clips as 150 frames and Section 4.2.1 states video durations of 150 to 300 frames. These numbers should be reconciled.
  4. [Section 5.3.1, Tables 3-6] The comparison with YOLO baselines is presented without stating the operating confidence threshold used for the main tables. Tables 5 and 6 show that YOLO variants produce near-zero results at confidence thresholds 0.25 and 0.10, and only become competitive at 0.01; the main comparisons in Tables 3 and 4 appear to use conf = 0.01. This makes the baseline comparison sensitive to threshold selection and should be stated and justified in the main text. In addition, Table 6 uses frame-level column headers ('±≤0', '±≤1', etc.) for pixel-level metrics, and the last column contains average-error values, which is a formatting error that should be corrected.
minor comments (5)
  1. [Eq. (10)] The formula for Err_ibf_avg lacks an absolute value; as written it is a signed mean difference, not an average error. Please use |tgt - tpred| inside the summation, and clarify the text in the caption of Table 3, which describes this metric as average accuracy across tolerance levels.
  2. [Section 5.4.1] The sentence 'the overall average error (Err_ibf_avg) deteriorates from 4.77 to 3.69' is backwards, since a lower value is better; the with-memory result of 3.69 is an improvement, not a deterioration.
  3. [Algorithm 1] In the Deployment branch, line 25 loops 'for t = tgt to T−1', but tgt is the ground-truth onset frame and is not available during deployment; this should be tb. There is also a typo on the initialization line: 'Pt−1← Ppred, , It−1← Itb' contains a doubled comma.
  4. [Section 5.2] The naming of CoTracker is inconsistent: the text uses 'C otrackerV3', 'CotrackerV3', and 'Co-tracker' in different places. Please unify the spelling.
  5. [Table 6] The last row of Table 6 reports Erravg values (e.g., 146.91) under the '±≤16' column header; these should be moved to a separate 'Erravg' column with appropriate pixel-error headers.

Circularity Check

1 steps flagged · score 3.0 of 10

Onset-detection benchmark is degenerate: since every test clip has ground-truth onset at frame 120, a constant predictor attains 100% under the paper's own metric, making the reported 96.85% SOTA uninterpretable; other stages are not circular.

  1. other [Section 3.3.1 (Bleeding Source Detection Dataset) and Section 5.1, Eq. (9)]
    "Each clip consists of 150 consecutive frames, and the initial bleeding source consistently occurs at frame 120. ... For validation and testing, the full clips are used without frame skipping to ensure consistent and reliable evaluation."

    Eq. (9) defines frame-level accuracy by comparing a predicted onset tpred with ground truth tgt. Section 3.3.1 fixes tgt = 120 for every test clip and states that testing uses full clips without frame skipping. Therefore the constant predictor tpred = 120 achieves Acc±k = 100% for all k >= 0 by construction. The paper reports only 49.92% at ±0 and 96.85% at ±8 (Table 3), i.e., below this trivial baseline, and never reports the constant baseline or the distribution of onset frames. The claimed state-of-the-art onset-detection result is thus not evidence of detecting a variable onset; the benchmark's accuracy metric reduces to a constant-target matching problem.

full rationale

Apart from the degenerate onset benchmark, the paper's derivation chain is not circular. The tracking pseudo-labels (Eqs. 6-7) are generated by an external pre-trained Track-On, XFeat feature matching, and Kalman filtering, not by the final fine-tuned model, and tracking evaluation is performed on held-out manual annotations. Detection and tracking evaluations explicitly use ground-truth initialization (Section 4.3, Algorithm 1), which limits end-to-end validity but is not a concealed circularity. Hyperparameters such as α=γ=0.5, N=60, r=50, and S>0.7 are hand-set rather than fitted values renamed as predictions. No load-bearing self-citation chain is present. The serious weakness is the constant-onset construction in Section 3.3.1 combined with Eq. (9): it makes frame-level onset accuracy meaningless as a measure of detection, directly undermining the first stage of the claimed workflow. The deployment evaluation (Section 6.4) is qualitative and based on clinician interviews, which is a stated limitation rather than a circular step.

Assumptions & free parameters 10 free parameters · 5 assumptions · 1 invented entities

The central claims rest on hand-set thresholds, the pseudo-label generation assumptions, and the single-point annotation model. These are all stated in the paper, but none are independently verified or provided with sensitivity analyses.

free parameters (10)
  • clean_view_threshold_alpha = 0.5
    RGB-HSV difference threshold for storing a 'clean view' frame in memory (Eq. 1); hand-set without sensitivity analysis.
  • clean_view_threshold_gamma = 0.5
    Optical flow difference threshold for clean-view selection (Eq. 1); hand-set.
  • sliding_window_size_N = 60
    Temporal window length for bleeding onset detection (Section 4.1.1); hand-set.
  • pseudo_label_offset_n = 10
    Number of frames ahead for optical-flow pseudo-label propagation in detection (Section 4.1.2); hand-set.
  • detection_loss_weights = lambda1=0.5, lambda2=0.5, delta=1
    Weights balancing real and pseudo heatmaps and coordinate regression in Eq. (5); hand-set.
  • pseudo_keypoint_radius_r = 50 pixels
    Radius around the annotated bleeding source for extracting matched keypoints (Eq. 6); hand-set.
  • matching_confidence_threshold_S = 0.7
    XFeat match confidence threshold for accepting pseudo-keypoints (Eq. 6); hand-set.
  • tracking_loss_weights = alpha1=0.6, alpha2=0.4
    Weights for ground-truth vs pseudo-label Huber losses in tracking (Eq. 8); hand-set.
  • onset_confidence_threshold = 0.5
    Confidence threshold for accepting a sliding-window onset prediction (Eq. 2); hand-set.
  • memory_refresh_interval = 60 frames
    Deployment strategy in Section 6.3 that refreshes tracking memory every N frames; N=60 chosen on the basis of observed drift.
assumptions (5)
  • domain assumption The bleeding source is a single, expert-annotatable point per frame, and expert consensus plus majority voting defines the ground truth.
    The entire dataset annotation pipeline in Section 3.2 assumes a unique bleeding-source point exists and is identifiable in each frame.
  • domain assumption Clips of 150 frames with the initial bleeding source at frame 120 represent real ESD bleeding events.
    Section 3.3.1 states this construction; if onset is artificially fixed, temporal onset detection becomes trivial and the reported accuracies are not meaningful for real-world variability.
  • ad hoc to paper Optical flow can propagate the bleeding source coordinate for up to n=10 frames with acceptable accuracy.
    Section 4.1.2 uses optical flow to create pseudo-labels for detection training; no validation of propagated pseudo-label accuracy is provided.
  • ad hoc to paper Trajectories from the off-the-shelf Track-On model, after Kalman smoothing with XFeat endpoint correction, yield pseudo-labels accurate enough to train the tracking model.
    Section 4.2.1 relies on these pseudo-labels for dense supervision; the paper does not measure pseudo-label error or its correlation with the test distribution.
  • domain assumption Pixel distance thresholds such as 100 px correspond to clinically acceptable localization error in the ESD view.
    Metrics in Section 5.1 adopt thresholds from prior point-tracking work without a clinical calibration argument for ESD.
invented entities (1)
  • No new physical or datalogical entities are introduced
    purpose: No new particles, forces, dimensions, or invented constructs are postulated; MDCFM and MDG are architectural modules rather than entities.
    The paper introduces only neural network modules and dataset labels, which are not entities in the sense this ledger tracks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BleedOrigin: Dynamic Bleeding Source Localization in Endoscopic Submucosal Dissection via Dual-Stage Detection and Tracking." pith.science (2026). https://pith.science/paper/YFCRD3DG

@misc{pith2026250715094,
  author       = {Pith},
  title        = {Pith review of: BleedOrigin: Dynamic Bleeding Source Localization in Endoscopic Submucosal Dissection via Dual-Stage Detection and Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YFCRD3DG}},
  note         = {Machine review of arXiv:2507.15094}
}
abstract

Intraoperative bleeding during Endoscopic Submucosal Dissection (ESD) poses significant risks, demanding precise, real-time localization and continuous monitoring of the bleeding source for effective hemostatic intervention. In particular, endoscopists have to repeatedly flush to clear blood, allowing only milliseconds to identify bleeding sources, an inefficient process that prolongs operations and elevates patient risks. However, current Artificial Intelligence (AI) methods primarily focus on bleeding region segmentation, overlooking the critical need for accurate bleeding source detection and temporal tracking in the challenging ESD environment, which is marked by frequent visual obstructions and dynamic scene changes. This gap is widened by the lack of specialized datasets, hindering the development of robust AI-assisted guidance systems. To address these challenges, we introduce BleedOrigin-Bench, the first comprehensive ESD bleeding source dataset, featuring 1,771 expert-annotated bleeding sources across 106,222 frames from 44 procedures, supplemented with 39,755 pseudo-labeled frames. This benchmark covers 8 anatomical sites and 6 challenging clinical scenarios. We also present BleedOrigin-Net, a novel dual-stage detection-tracking framework for the bleeding source localization in ESD procedures, addressing the complete workflow from bleeding onset detection to continuous spatial tracking. We compare with widely-used object detection models (YOLOv11/v12), multimodal large language models, and point tracking methods. Extensive evaluation demonstrates state-of-the-art performance, achieving 96.85% frame-level accuracy ($\pm\leq8$ frames) for bleeding onset detection, 70.24% pixel-level accuracy ($\leq100$ px) for initial source detection, and 96.11% pixel-level accuracy ($\leq100$ px) for point tracking.

Figures

Figures reproduced from arXiv: 2507.15094 by the authors.

Figure 1
Figure 1. Overview of the Motivation and Methodology. (A) Motivation: persistent bleeding obscures the surgical field, necessitating repeated flushing to achieve [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our BleedOrigin-Bench dataset. We selected 485 bleeding video clips from 44 patients for analysis, processed in two stages: (1) initial [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Overview of the dataset annotation pipeline. Our multi-stage annotation pipeline is used to create the high-quality BleedOrigin-Bench dataset. Raw videos [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Dataset statistics by patient. (Top) Distribution of video clips per patient across train(yellow), validation(blue), and test(green) splits. (Bottom) The [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Overview of the proposed BleedOrigin-Detect for detecting the initial bleeding frame and point in ESD procedures. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the BleedOrigin-Track model. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: (A) Illustration of pixel-level distances overlaid on real endoscopic frames, highlighting the spatial scale of deviations (e.g., 10px to 100px). (B) Temporal [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Comparison of MLLMs’ predictions for bleeding source localization in intraoperative endoscopic images. (A) Task prompt and visual guidance provided [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison of bleeding source tracking across six di [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Visual comparison of attention maps from di [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Visualization of our detection and tracking results on full-length ESD videos. Frame indices are shown in the top-right corners. In (A) and (B), the [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of bleeding source tracking with di [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Frame-wise pixel error of bleeding source tracking under di [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: The scene complexity progresses through three distinct tiers: low di [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 60 canonical work pages

  1. [1]

    Claude: An AI Assistant by Anthropic

    Anthropic, 2024. Claude: An AI Assistant by Anthropic. https://www. anthropic.com/claude. Accessed: 27 Jun 2024

    work page 2024

  2. [2]

    Track-On: Transformer-based Online Point Tracking with Memory

    Aydemir, G., Cai, X., Xie, W., G ¨uney, F., 2025. Track-on: Transformer-based online point tracking with memory. arXiv preprint arXiv:2501.18487

    work page Pith review arXiv 2025

  3. [3]

    Qwen2.5-vl technical report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y ., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y ., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J., 2025. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  4. [4]

    Automated bleeding identification in surgical videos using deep learning

    Bamba, Y ., Ogawa, S., Itabashi, M., Kameoka, S., Okamoto, T., Ya- mamoto, M., Yamaguchi, S., 2022. Automated bleeding identification in surgical videos using deep learning. Tokyo Women’s Medical University Journal 6, 117–125

    work page 2022

  5. [5]

    Pattern recognition and machine learning

    Bishop, C.M., Nasrabadi, N.M., 2006. Pattern recognition and machine learning. volume 4. Springer

    work page 2006

  6. [6]

    Perception encoder: The best visual embeddings are not at the output of the network

    Bolya, D., Huang, P.Y ., Sun, P., Cho, J.H., Madotto, A., Wei, C., Ma, T., Zhi, J., Rajasegaran, J., Rasheed, H., Wang, J., Monteiro, M., Xu, H., Dong, S., Ravi, N., Li, D., Doll´ar, P., Feichtenhofer, C., 2025. Perception encoder: The best visual embeddings are not at the output of the network. arXiv:2504.13181

    arXiv 2025

  7. [7]

    Visual features for im- proving endoscopic bleeding detection using convolutional neural net- works

    Brzeski, A., Dziubich, T., Krawczyk, H., 2023. Visual features for im- proving endoscopic bleeding detection using convolutional neural net- works. Sensors 23, 9717

    work page 2023

  8. [8]

    Cao, J., Yip, H.C., Chen, Y ., Scheppach, M., Luo, X., Yang, H., Cheng, M.K., Long, Y ., Jin, Y ., Chiu, P.W.Y ., Yam, Y ., Meng, H.M.L., Dou, Q.,

    work page

Show all 79 references
  1. [9]

    Chen, T., Lyu, Q., Bai, L., Guo, E., Gao, H., Yang, X., Ren, H., Zhou, L.,

  2. [10]

    Tempura: Temporal event masked prediction and understanding for reasoning in action

    Cheng, J.H., Wang, V ., Wang, H., Zhou, H., Peng, Y .H., Liu, H.I., Huang, H.W., Chen, K.M., Yang, C.Y ., Chai, W., Chen, Y .L., Vineet, V ., Cai, Q., Jenq-Neng, H., 2025. Tempura: Temporal event masked prediction and understanding for reasoning in action. arXiv preprint arXiv...

    2025 arXiv

  3. [11]

    Boot- sTAP: Bootstrapped training for tracking-any-point

    Doersch, C., Luc, P., Yang, Y ., Gokay, D., Koppula, S., Gupta, A., Hey- ward, J., Rocco, I., Goroshin, R., Carreira, J., Zisserman, A., 2024. Boot- sTAP: Bootstrapped training for tracking-any-point. Asian Conference on Computer Vision

    2024

  4. [12]

    TAPIR: Tracking any point with per-frame initialization and temporal refinement, in: Proceedings of the IEEE /CVF International Conference on Computer Vision, pp

    Doersch, C., Yang, Y ., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y ., Car- reira, J., Zisserman, A., 2023. TAPIR: Tracking any point with per-frame initialization and temporal refinement, in: Proceedings of the IEEE /CVF International Conference on Computer Vision, pp. 10061–10072

    2023

  5. [13]

    Parameter-efficient fine-tuning for medical image analysis: The missed opportunity

    Dutt, R., Ericsson, L., Sanchez, P., Tsaftaris, S.A., Hospedales, T., 2023. Parameter-efficient fine-tuning for medical image analysis: The missed opportunity. arXiv preprint arXiv:2305.08252

    2023 arXiv

  6. [14]

    Development of a novel endoscopic hemostasis-assisted navigation ai system in the standardization of post- esd coagulation

    Fujinami, H., Kuraishi, S., Teramoto, A., Shimada, S., Takahashi, S., Ando, T., Yasuda, I., 2024. Development of a novel endoscopic hemostasis-assisted navigation ai system in the standardization of post- esd coagulation. Endoscopy International Open 12, E520–E525

    2024

  7. [15]

    Automatic detection of surgical haemorrhage using computer vision

    Garcia-Martinez, A., Vicente-Samper, J.M., Sabater-Navarro, J.M., 2017. Automatic detection of surgical haemorrhage using computer vision. Ar- tificial intelligence in medicine 78, 55–60

    2017

  8. [16]

    Gemini: a family of highly capable multimodal models

    Gemini Team, G., 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2023 arXiv

  9. [17]

    Generative adversarial net- works

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y ., 2020. Generative adversarial net- works. Communications of the ACM 63, 139–144

    2020

  10. [18]

    Kubric: a scalable dataset generator

    Gre ff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y ., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., Kipf, T., Kundu, A., Lagun, D., Laradji, I., Liu, H.T.D., Meyer, H., Miao, Y ., Nowrouzezahrai, D., Oztireli, C., Pot, E., Radwan, N., Rebain, D....

    2022

  11. [19]

    Kalman filtering: Theory and Prac- tice with MATLAB

    Grewal, M.S., Andrews, A.P., 2014. Kalman filtering: Theory and Prac- tice with MATLAB. John Wiley & Sons

    2014

  12. [20]

    Parameter- efficient fine-tuning for large models: A comprehensive survey

    Han, Z., Gao, C., Liu, J., Zhang, J., Zhang, S.Q., 2024. Parameter- efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608

    2024 arXiv

  13. [21]

    Wcebleedgen: A wireless capsule endoscopy dataset and its benchmarking for automatic bleeding classification, detection, and segmentation

    Handa, P., Dhir, M., Mahbod, A., Schwarzhans, F., Woitek, R., Goel, N., Gunjan, D., 2024. Wcebleedgen: A wireless capsule endoscopy dataset and its benchmarking for automatic bleeding classification, detection, and segmentation. arXiv preprint arXiv:2408.12466

    2024 arXiv

  14. [22]

    Real-time detection of active bleeding in la- paroscopic colectomy using artificial intelligence

    Horita, K., Hida, K., Itatani, Y ., Fujita, H., Hidaka, Y ., Yamamoto, G., Ito, M., Obama, K., 2024. Real-time detection of active bleeding in la- paroscopic colectomy using artificial intelligence. Surgical Endoscopy 38, 3461–3469

    2024

  15. [23]

    Lora: Low-rank adaptation of large language models

    Hu, E.J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., Chen, W., 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 3

    2022

  16. [24]

    Automatic bleeding detection in laparoscopic surgery based on a faster region-based convolutional neural network

    Hua, S., Gao, J., Wang, Z., Yeerkenbieke, P., Li, J., Wang, J., He, G., Jiang, J., Lu, Y ., Yu, Q., Han, X., Liao, Q., Wu, W., 2022. Automatic bleeding detection in laparoscopic surgery based on a faster region-based convolutional neural network. Annals of Translational Medici...

    2022

  17. [25]

    Robust estimation of a location parameter

    Huber, P., 1992. Robust estimation of a location parameter. breakthroughs in statistics: Methodology and distribution

    1992

  18. [26]

    Cotracker3: Simpler and better point tracking by pseudo- labelling real videos

    Karaev, N., Makarov, I., Wang, J., Neverova, N., Vedaldi, A., Rupprecht, C., 2024a. Cotracker3: Simpler and better point tracking by pseudo- labelling real videos. arXiv preprint arXiv:2410.11831

    arXiv

  19. [27]

    Cotracker: It is better to track together, in: European Confer- ence on Computer Vision, Springer

    Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C., 2024b. Cotracker: It is better to track together, in: European Confer- ence on Computer Vision, Springer. pp. 18–35

  20. [28]

    Yolov11: An overview of the key archi- tectural enhancements

    Khanam, R., Hussain, M., 2024. Yolov11: An overview of the key archi- tectural enhancements. arXiv preprint arXiv:2410.17725

    2024 arXiv

  21. [29]

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y ., Doll ´ar, P., Girshick, R.,

  22. [30]

    Tapvid-3d: A benchmark for tracking any point in 3d

    Koppula, S., Rocco, I., Yang, Y ., Heyward, J., Carreira, J., Zisserman, A., Brostow, G., Doersch, C., 2024. Tapvid-3d: A benchmark for tracking any point in 3d. arXiv preprint arXiv:2407.05921

    2024 arXiv

  23. [31]

    Semi-supervised surgical video semantic segmentation with cross supervision of inter-frame, in: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), IEEE

    Li, D., Hu, Y ., Shen, J., Hao, L., Liu, J., 2023a. Semi-supervised surgical video semantic segmentation with cross supervision of inter-frame, in: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), IEEE. pp. 1–5

    2023

  24. [32]

    4015–4026

    Segment anything, in: Proceedings of the IEEE /CVF international conference on computer vision, pp. 4015–4026

  25. [33]

    Semivt-surge: Semi- supervised video transformer for surgical phase recognition

    Li, Y ., de Jong, R., Nasirihaghighi, S., Jaspers, T., van Jaarsveld, R., Kuiper, G., van Hillegersberg, R., van der Sommen, F., Ru- urda, J., Breeuwer, M., Al Khalil, Y ., 2025. Semivt-surge: Semi- supervised video transformer for surgical phase recognition. arXiv preprint ar...

    2025 arXiv

  26. [34]

    Vision foundation models in medical image analysis: Advances and chal- lenges

    Liang, P., Pu, B., Huang, H., Li, Y ., Wang, H., Ma, W., Chang, Q., 2025. Vision foundation models in medical image analysis: Advances and chal- lenges. arXiv preprint arXiv:2502.14584

    2025 arXiv

  27. [35]

    Li, J., Jin, Y ., Chen, Y ., Yip, H.C., Scheppach, M., Chiu, P.W.Y ., Yam, Y ., Meng, H.M.L., Dou, Q., 2023b. Imitation learning from expert video data for dissection trajectory prediction in endoscopic surgical procedure, in: International Conference on Medical Image Computin...

  28. [36]

    Research on detection method of bleeding point in two-dimensional dsa image based on para- metric color imaging

    Liu, J., Zhou, X., Guan, W., Gong, S., Liu, J., 2022. Research on detection method of bleeding point in two-dimensional dsa image based on para- metric color imaging. Computers in Biology and Medicine 146, 105496

    2022

  29. [37]

    Looking fast and slow: Memory-guided mobile video object detection

    Liu, M., Zhu, M., White, M., Li, Y ., Kalenichenko, D., 2019. Looking fast and slow: Memory-guided mobile video object detection. arXiv preprint arXiv:1903.10172

    2019 arXiv

  30. [38]

    Focal loss for dense object detection, in: Proceedings of the IEEE international confer- ence on computer vision, pp

    Lin, T.Y ., Goyal, P., Girshick, R., He, K., Doll´ar, P., 2017. Focal loss for dense object detection, in: Proceedings of the IEEE international confer- ence on computer vision, pp. 2980–2988

    2017

  31. [39]

    Hemoset: The first blood segmentation dataset for automation of hemostasis management, in: 2024 International Sym- posium on Medical Robotics (ISMR), IEEE

    Miao, A.J., Lin, S., Lu, J., Richter, F., Ostrander, B., Funk, E.K., Orosco, R.K., Yip, M.C., 2024. Hemoset: The first blood segmentation dataset for automation of hemostasis management, in: 2024 International Sym- posium on Medical Robotics (ISMR), IEEE. pp. 1–7

    2024

  32. [40]

    Cross- modal transformer-based streaming dense video captioning with neural ode temporal localization

    Muksimova, S., Umirzakova, S., Sultanov, M., Im Cho, Y ., 2025. Cross- modal transformer-based streaming dense video captioning with neural ode temporal localization. Sensors (Basel, Switzerland) 25, 707

    2025

  33. [41]

    Current advancements in application of artificial intelligence in clinical decision-making by gastroenterologists in gastrointestinal bleeding

    Maulahela, H., Annisa, N.G., 2022. Current advancements in application of artificial intelligence in clinical decision-making by gastroenterologists in gastrointestinal bleeding. Artificial Intelligence in Gastroenterology 3, 13–20

    2022

  34. [42]

    A review of applica- tion of deep learning in endoscopic image processing

    Nie, Z., Xu, M., Wang, Z., Lu, X., Song, W., 2024. A review of applica- tion of deep learning in endoscopic image processing. Journal of Imaging 10, 275

    2024

  35. [43]

    A narrative review of foundation models for medical image segmentation: Zero-shot performance evaluation on di- verse modalities

    Noh, S., Lee, B.D., 2025. A narrative review of foundation models for medical image segmentation: Zero-shot performance evaluation on di- verse modalities. Quantitative Imaging in Medicine and Surgery

    2025

  36. [44]

    Mft: Long-term tracking of every pixel, in: Proceedings of the IEEE /CVF Winter Conference on Applica- tions of Computer Vision, pp

    Neoral, M., ˇSer`ych, J., Matas, J., 2024. Mft: Long-term tracking of every pixel, in: Proceedings of the IEEE /CVF Winter Conference on Applica- tions of Computer Vision, pp. 6837–6847

    2024

  37. [45]

    Dinov2: Learning robust visual fea- tures without supervision

    Oquab, M., Darcet, T., Moutakanni, T., V o, H.V ., Szafraniec, M., Khali- dov, V ., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y ., Xu, H., Sharma, V ., Li, S.W., Galuba, W., Rabbat, M., Ass- ran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H.,...

    2023

  38. [46]

    Autonomous blood suction for robot-assisted surgery: A sim-to-real reinforcement learning approach

    Ou, Y ., Soleymani, A., Li, X., Tavakoli, M., 2024. Autonomous blood suction for robot-assisted surgery: A sim-to-real reinforcement learning approach. IEEE Robotics and Automation Letters

    2024

  39. [47]

    Gpt-4 technical report

    OpenAI, 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    2023 arXiv

  40. [48]

    Synergistic bleeding region and point detection in surgical videos

    Pei, J., Zhou, Z., Guo, D., Li, Z., Qin, J., Du, B., Heng, P.A., 2025. Synergistic bleeding region and point detection in surgical videos. arXiv preprint arXiv:2503.22174

    2025

  41. [49]

    Xfeat: Accelerated features for lightweight image matching, in: Pro- ceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp

    Potje, G., Cadar, F., Araujo, A., Martins, R., Nascimento, E.R., 2024. Xfeat: Accelerated features for lightweight image matching, in: Pro- ceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp. 2682–2691

    2024

  42. [50]

    Learning autonomous surgical irrigation and suction with the da vinci research kit using reinforcement learning

    Ou, Y ., Tavakoli, M., 2025. Learning autonomous surgical irrigation and suction with the da vinci research kit using reinforcement learning. IEEE Transactions on Automation Science and Engineering

    2025

  43. [51]

    Video- based computer-aided laparoscopic bleeding management: a space-time memory neural network with positional encoding and adversarial domain adaptation., in: MIDL, pp

    Rabbani, N., S `eve-d’Erceville, C., Bourdel, N., Bartoli, A., 2022. Video- based computer-aided laparoscopic bleeding management: a space-time memory neural network with positional encoding and adversarial domain adaptation., in: MIDL, pp. 961–974

    2022

  44. [52]

    An entropy-based approach to detect and localize intraoperative bleeding during minimally invasive surgery

    Rahbar, M.D., Reisner, L., Ying, H., Pandya, A., 2020. An entropy-based approach to detect and localize intraoperative bleeding during minimally invasive surgery. The International Journal of Medical Robotics and Com- puter Assisted Surgery 16, 1–9

    2020

  45. [53]

    Adapting vision-language foundation model for next generation medical ultrasound image analysis

    Qu, J., Han, X., Xiao, T., Ai, J., Wu, J., Zhao, T., Qin, J., King, A.D., Chu, W.C.W., Cai, J., Ying, M.T.C., 2025. Adapting vision-language foundation model for next generation medical ultrasound image analysis. arXiv preprint arXiv:2506.08849 . 26

    2025 arXiv

  46. [54]

    Faster r-cnn: Towards real- time object detection with region proposal networks

    Ren, S., He, K., Girshick, R., Sun, J., 2015. Faster r-cnn: Towards real- time object detection with region proposal networks. Advances in neural information processing systems 28

    2015

  47. [55]

    Autonomous robotic suction to clear the surgical field for hemostasis using image-based blood flow detection

    Richter, F., Shen, S., Liu, F., Huang, J., Funk, E.K., Orosco, R.K., Yip, M.C., 2021. Autonomous robotic suction to clear the surgical field for hemostasis using image-based blood flow detection. IEEE Robotics and Automation Letters 6, 1383–1390

    2021

  48. [56]

    You only look once: Unified, real-time object detection, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Redmon, J., Divvala, S., Girshick, R., Farhadi, A., 2016. You only look once: Unified, real-time object detection, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 779–788

    2016

  49. [57]

    Complications related to gastric endoscopic submucosal dissection and their managements

    Saito, I., Tsuji, Y ., Sakaguchi, Y ., Niimi, K., Ono, S., Kodashima, S., Ya- mamichi, N., Fujishiro, M., Koike, K., 2014. Complications related to gastric endoscopic submucosal dissection and their managements. Clini- cal endoscopy 47, 398–403

    2014

  50. [58]

    Softed: Metrics for soft evaluation of time series event detection

    Salles, R., Lima, J., Reis, M., Coutinho, R., Pacitti, E., Masseglia, F., Ak- barinia, R., Chen, C., Garibaldi, J., Porto, F., Ogasawara, E., 2024. Softed: Metrics for soft evaluation of time series event detection. Computers & Industrial Engineering 198, 110728

    2024

  51. [59]

    Deep learning-based lesion detection in endoscopy: A systematic literature review

    Rochmawati, N., Fatichah, C., Amaliah, B., Raharjo, A.B., Dumont, F., Thibaudeau, E., Dumas, C., 2025. Deep learning-based lesion detection in endoscopy: A systematic literature review. IEEE Access

    2025

  52. [60]

    Surgical tattoos in infrared: A dataset for quantifying tissue tracking and mapping

    Schmidt, A., Mohareri, O., DiMaio, S., Salcudean, S.E., 2024. Surgical tattoos in infrared: A dataset for quantifying tissue tracking and mapping. IEEE Transactions on Medical Imaging

    2024

  53. [61]

    Mftiq: Multi-flow tracker with in- dependent matching quality estimation, in: 2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (W ACV), IEEE

    Serych, J., Neoral, M., Matas, J., 2025. Mftiq: Multi-flow tracker with in- dependent matching quality estimation, in: 2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (W ACV), IEEE. pp. 8079– 8089

    2025

  54. [62]

    Point tracking in surgery–the 2024 surgi- cal tattoos in infrared (stir) challenge

    Schmidt, A., colleagues, 2025. Point tracking in surgery–the 2024 surgi- cal tattoos in infrared (stir) challenge. arXiv preprint arXiv:2503.24306

    2025 arXiv

  55. [63]

    Bleeding alert map (bam): The identification method of the bleeding source in real organs using datasets made on mimicking organs

    Sogabe, M., Ishikawa, K., Takamatsu, T., Takeuchi, K., Kanno, T., Fu- jimoto, K., Miyazaki, T., Kawase, T., Sato, T., Kawashima, K., 2023. Bleeding alert map (bam): The identification method of the bleeding source in real organs using datasets made on mimicking organs. Array 1...

    2023

  56. [64]

    Online temporal action localization with memory-augmented transformer, in: European Confer- ence on Computer Vision, Springer

    Song, Y ., Kim, D., Cho, M., Kwak, S., 2024. Online temporal action localization with memory-augmented transformer, in: European Confer- ence on Computer Vision, Springer. pp. 74–91

    2024

  57. [65]

    Semi-supervised learning with progressive unlabeled data excavation for label-e fficient surgical work- flow recognition

    Shi, X., Jin, Y ., Dou, Q., Heng, P.A., 2021. Semi-supervised learning with progressive unlabeled data excavation for label-e fficient surgical work- flow recognition. Medical Image Analysis 73, 102158

    2021

  58. [66]

    Yolov12: Attention-centric real- time object detectors

    Tian, Y ., Ye, Q., Doermann, D., 2025. Yolov12: Attention-centric real- time object detectors. arXiv preprint arXiv:2502.12524

    2025 arXiv

  59. [67]

    Endochat: Grounded multimodal large language model for endoscopic surgery

    Wang, G., Bai, L., Wang, J., Yuan, K., Li, Z., Jiang, T., He, X., Wu, J., Chen, Z., Lei, Z., Liu, H., Wang, J., Zhang, F., Padoy, N., Navab, N., Ren, H., 2025a. Endochat: Grounded multimodal large language model for endoscopic surgery. arXiv preprint arXiv:2501.11347

    arXiv

  60. [68]

    Motion2vec: Semi-supervised representation learning from surgical videos, in: 2020 IEEE International Conference on Robotics and Automation (ICRA), IEEE

    Tanwani, A.K., Sermanet, P., Yan, A., Anand, R., Phielipp, M., Goldberg, K., 2020. Motion2vec: Semi-supervised representation learning from surgical videos, in: 2020 IEEE International Conference on Robotics and Automation (ICRA), IEEE. pp. 2174–2181

    2020

  61. [69]

    Video object detection with an aligned spatial- temporal memory, in: Proceedings of the European conference on com- puter vision (ECCV), pp

    Xiao, F., Lee, Y .J., 2018. Video object detection with an aligned spatial- temporal memory, in: Proceedings of the European conference on com- puter vision (ECCV), pp. 485–501

    2018

  62. [70]

    Artificial intelligence assisted to- pographic mapping system for endoscopic submucosal dissection speci- mens

    Xiao, Y ., Song, Z., Zou, S., You, Y ., Cui, J., Wang, S., Ku, C., Wu, X., Xue, X., Han, W., Zhou, W., 2022. Artificial intelligence assisted to- pographic mapping system for endoscopic submucosal dissection speci- mens. Frontiers in Medicine 9, 822731

    2022

  63. [71]

    Wang, X., Qiao, P., Dou, Y ., 2025b. Hemo-fs-sam2: Few-shot tuning on hemorrhage segmentation of segment anything model 2, in: 2025 8th In- ternational Conference on Advanced Algorithms and Control Engineering (ICAACE), IEEE. pp. 1079–1082

    2025

  64. [72]

    Pdzseg: adapting the foundation model for dissec- tion zone segmentation with visual prompts in robot-assisted endoscopic submucosal dissection

    Xu, M., Mo, W., Wang, G., Gao, H., Wang, A., Zhong, N., Li, Z., Yang, X., Ren, H., 2025. Pdzseg: adapting the foundation model for dissec- tion zone segmentation with visual prompts in robot-assisted endoscopic submucosal dissection. International Journal of Computer Assisted ...

    2025

  65. [73]

    Samu- rai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory

    Yang, C.Y ., Huang, H.W., Chai, W., Jiang, Z., Hwang, J.N., 2024. Samu- rai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory. arXiv preprint arXiv:2411.11922

    2024 arXiv

  66. [74]

    Etsm: Automating dissection trajectory suggestion and confidence map-based safety margin prediction for robot-assisted en- doscopic submucosal dissection

    Xu, M., Mo, W., Wang, G., Gao, H., Wang, A., Bai, L., Lyu, C., Yang, X., Li, Z., Ren, H., 2024. Etsm: Automating dissection trajectory suggestion and confidence map-based safety margin prediction for robot-assisted en- doscopic submucosal dissection. arXiv preprint arXiv:2411.18884

    2024 arXiv

  67. [75]

    Automatic bleeding frame detection in the wireless capsule endoscopy images, in: 2015 IEEE International Con- ference on Robotics and Automation (ICRA), IEEE

    Yuan, Y ., Meng, M.Q.H., 2015. Automatic bleeding frame detection in the wireless capsule endoscopy images, in: 2015 IEEE International Con- ference on Robotics and Automation (ICRA), IEEE. pp. 1310–1315

    2015

  68. [76]

    Adalora: Adaptive budget allocation for parameter-efficient fine-tuning

    Zhang, Q., Chen, M., Bukharin, A., Karampatziakis, N., He, P., Cheng, Y ., Chen, W., Zhao, T., 2023. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512 . 27

    2023 arXiv

  69. [77]

    Pace: A large-scale dataset with pose annotations in cluttered environments, in: European Conference on Computer Vision, Springer

    You, Y ., Xiong, K., Yang, Z., Huang, Z., Zhou, J., Shi, R., Fang, Z., Harley, A.W., Guibas, L., Lu, C., 2024. Pace: A large-scale dataset with pose annotations in cluttered environments, in: European Conference on Computer Vision, Springer. pp. 473–489

    2024

  70. [2023]

    Nature Communications 14, 6676

    Intelligent surgical workflow recognition for endoscopic submu- cosal dissection with real-time animal study. Nature Communications 14, 6676

  71. [2024]

    Lightdi ff: Surgical endoscopic image low-light enhancement with t-diffusion, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 369–379. 25

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.