Abstract

Medical imaging traditionally assumes that high-fidelity, lossless data are required for reliable computer-aided diagnosis (CAD), an assumption inherited by AI models trained on pristine DICOM volumes. While 2D radiographs have shown tolerance to smartphone recapture, it remains unknown whether 3D volumetric diagnosis can survive the severe degradation introduced when a camera records slice scrolling as a video stream. This question is increasingly relevant as DICOM-level access and PACS integration impose substantial infrastructural barriers, particularly in resource-limited settings. We introduce ScreenCAD, a framework designed to test whether a 3D medical volume, transformed into a camera-captured video, still preserves sufficient semantic information for AI diagnosis. ScreenCAD combines cascaded ROI extraction, a frozen medical foundation model with strong semantic invariance, and a Task-Token Mixture-of-Experts aggregator to operate directly on low-fidelity, variable-length video streams. On CT-RATE with real-world screen recordings, ScreenCAD achieves competitive diagnostic performance under video inputs, while conventional 3D baselines degrade sharply. These results provide initial evidence that volumetric diagnosis can remain feasible under substantial information loss, highlighting camera-based acquisition as a promising interface for future medical AI systems when direct PACS integration is infeasible. Our code will be released at GitHub.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4759_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WuJun_Diagnosing_MICCAI2026,
        author = { Wu, Junlei AND Huang, Haolin AND Li, Jiaming AND Wang, Qian},
        title = { { Diagnosing 3D Volumes from Camera Video: A Semantic-Robust Framework Without Direct DICOM Access } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper describes the actual diagnosis procedure performed by a radiologist or other clinician - by scrolling through high resolution tomography images (here mainly CT) to cognitively forma 3D volume and asks the question on whether a camera captured scroll recording of the images as a video stream can actually create a useful information despite the many issues. It challenges the generally accepted image requirements and proves that this procedure is quite resilient with respect to diagnostic accuracy while the 3D model is severely degraded. The motivation is fascinating and actually quite in line with actual use of imaging data. This observation by itself is a great contribution.

  • 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.

    I uses a standard imaging data set and uses an interesting and well though through mix of different encoders, transformers, and task-aware algorithmic experts. It is very well written and employs a good methodology and experimental validation.

  • 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.

    Ultrasound 2D imaging moving along an (for example) arterial structure with the images continuously displayed could be a great application with an HDMI output directly fed into a recorder. The limitation to the CT dataset and a short discussion of other areas of application are the main limitations.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    A rare paper that actually looks at the use case and comes up with an innovation idea based on that.

  • 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?

    Novel approach, good methodology, applicable for other use cases, and well written.

  • 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 rebuttal resolves the baseline fairness concern. Since all models received identical preprocessed video inputs, the collapse of the 3D baselines stems entirely from their rigid architectural format constraints, which fail to handle variable lengths, pauses, or scrolling direction changes. Even with direct video fine-tuning, the baselines lag far behind ScreenCAD, and the ablation studies confirm that this success is driven by the newly designed aggregator rather than just the backbone.



Review #2

  • Please describe the contribution of the paper

    The authors propose a methodology to enable camera-captured volumetric medical imaging be usable for standard image analysis tasks, where the data is captured as a video (over the slices) with a camera recording the screen where the images are being displayed. The proposed method learns volume-level representations from the video stream and is robust to capture defects such as jitter, frame skip, camera artifacts, variable scrolling speed etc.

  • 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.
    • The paper is in general well written, and solves a problem: how to make camera captured data usable in a clinical context without direct dicom access, when this data is 3D.

    • The experimental set-up is appropriate and covers a wide range of scenarios.

    • The results suggest that their proposed approach can solve the problem

  • 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.
    • The main issue with this paper is that the comparison to the baselines is unfair. The video stream is fed to the competing methods (CT-Net, CT-Clip) as they are (trained on volumes), while their proposed method is fine-tuned to captured videos. Obviously the other methods are going to fail catastrophically; what would be interesting is to fine-tune the competing methods to take the videos as input (without any of the new ideas proposed in this paper) and then see what is the benefit of the proposed method. Without this, it is impossible to understand if the proposed method works for all the clever architectural choices in the preprocessor, or simply because it is fine-tuned to the data as it will be acquired. For instance, how would the proposed method perform if the actual dicoms were fed to it? most likely would also fail catastrophically!

    • The results are lacking depth: only average values are provided. some estimate of the dispersion is required, eg. variance, or even better confidence intervals. This would connect with some kind of statistical test on the significance of the numbers provided.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?
    • it would be good to comment on ethical or legal aspects of recording the screen with a camera, and using the recorded data for diagnostic purposes.
  • 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?

    The methos seems promising, but evaluation is flawed. If authors can clarify why it is not the case, then I am happy to reconsider my recommendation.

  • 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.

    Authors addressed my main concern (fairness of the baseline) by saying that the baseline methods are actually finetuned to the new type of data. I would have wished for more detail on this, and I hope this is included in the paper, but as it stands I lean more towards acceptance now.



Review #3

  • Please describe the contribution of the paper

    This paper address a novelty question whether a 3D CT volume, captured as a camera recording of a clinician scrolling through slices, retains enough information for AI diagnosis. The proposed system, ScreenCAD, answers this by chaining three components: a two-stage localization module to crop the medical viewport from raw video frames, a frozen MedDINOv3 encoder for slice-level feature extraction, and a Task-Token Mixture-of-Experts aggregator for multi-task classification. Experiments on CT-RATE show that ScreenCAD maintains near-DICOM performance on Pleural Effusion and Medical Material detection under camera-induced degradation, while standard 3D baselines collapse in the zero-shot video setting.

  • 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.

    The core research question is original and practically motivated. Removing the dependency on DICOM access and PACS infrastructure could genuinely broaden AI deployment in resource-limited clinical settings. The zero-shot Pleural Effusion result, where video AUC matches clean DICOM at 0.88 without any video fine-tuning, is the paper’s most compelling finding and provides real scientific insight about how low-frequency pathologies are encoded in semantic features. The ablation study shows that inter-slice modeling via the Transformer block is the most critical architectural component.

  • 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.ScreenCAD’s performances largely hinges on MedDINOv3.This is shown by the performance of ABMIL and Task Token compared to baselines. While ABMIL, replacing MedDINOv3 embeddings by Attention-Based MIL [9], perform worst than baselines, Task Token’s are better than baselines. The current evidence suggests ScreenCAD works because MedDINOv3 works.

    2.Baseline comparison is structurally unfair. CT-Net and CT-CLIP are 3D DICOM architectures with no video-handling capability whatsoever. Their zero-shot video collapse is expected and uninformative. On clean DICOM and fine-tuned video settings, these baselines are competitive with or superior to ScreenCAD in several metrics, which the paper underemphasizes. A fair comparison requires either equipping baselines with the same SVL preprocessing and MIL aggregation, or including video-native models such as VideoMAE as proper competitors.

    3.Evaluation is too narrow. While the video data collection setting is dataset-agnostic and the title claims “3D volumetric diagnosis” broadly, all results come from one dataset (CT-RATE) and two tasks. The baseline scores on CT-RATE are partially already reported in the original CT-RATE paper, limiting the independent experimental contribution. The title claims broad volumetric diagnosis but no MRI or PET modality is tested, and no additional CT pathologies are explored.

    4.The setting is less realistic than presented, and the data collection agent is unevaluated. Camera capture of 2D radiographs is practical because a single photograph suffices. Recording a clinician scrolling through hundreds of CT slices is a fundamentally different and more operationally demanding scenario. The authors acknowledge the controlled setup (single camera, single monitor, single viewer, ±15° angle range) but do not test generalization beyond it. More importantly, all training and test videos are generated by an automated agent described as following a “human-like browsing pattern.” Since this agent defines the entire data distribution, its scrolling statistics should be characterized quantitatively and validated against real human browsing behavior. Currently, the model may be learning to exploit the agent’s specific scrolling rhythm rather than generalizing to real clinical use.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Please see weakness items.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

Thanks for the valuable comments. We appreciate the recognition of our work and have addressed each of the reviewers’ concerns as follows. 1.Task Scope, Applications, and Generalization (R1, R4): We agree that ultrasound video is promising. HDMI capture removes degradation, while our work studies the complementary question of whether diagnosis remains reliable after severe ``analog-to-digital’’ camera degradation. Camera capture also avoids HDMI/PACS integration, which is useful in resource-limited settings. We will clarify this trade-off and discuss ultrasound, broader modalities, datasets, devices, viewers, and workflows as future validation. CT-RATE with Medical Material and Pleural Effusion is intended as a first reproducible testbed for this new task. 2.Baseline Protocol and Fairness (R2, R4): We will revise the manuscript to explicitly describe how video inputs are adapted for all baselines. All methods are evaluated after the same SVL preprocessing, so they receive the same localized medical viewport sequence. The remaining difference comes from each architecture’s intrinsic input constraints: CT-Net and CT-CLIP require fixed-size 3D tensors. For video inputs, we resize the SVL-extracted sequence to 420x420x402 for CT-Net. For CT-CLIP, since many videos contain more frames than its fixed depth, we first temporally sample frames to preserve coverage across the browsing trajectory, and then center-crop or pad the sampled volume to 480x480x240 following its original protocol. ScreenCAD’s variable-length MIL aggregation is not an extra preprocessing advantage, but a core architectural requirement for screen-recorded browsing videos, whose lengths vary due to scrolling speed, pauses, and reversals. We also fine-tuned CT-Net/CT-CLIP on captured videos; despite this, on Pleural Effusion Video FT they reach 0.71–0.73 AUC, while ScreenCAD reaches 0.90.3.MedDINOv3 Dependency and Architectural Contribution (R4): We agree that MedDINOv3 is an important component of ScreenCAD. Our point is that the backbone alone is insufficient. Table 1/2 compare ABMIL and Task Token under the same MedDINOv3 backbone, and Table 3 further ablates the aggregation components. For example, on Medical Material under Video ZS, replacing our Task-Token aggregator with standard ABMIL drops AUC from 0.72 to 0.58.This suggests that MedDINOv3 provides a strong clean-domain representation, while the proposed inter-slice and task-aware aggregation design provides additional robustness for low-quality screen-captured inputs. 4.Clean Performance, Metrics, and Statistical Analysis (R2, R4): We acknowledge that CT-Net/CT-CLIP are competitive on native DICOM inputs and may be slightly better on some secondary metrics in the Clean setting, which is expected because they are designed for native 3D DICOM CT volumes. Importantly, ScreenCAD remains competitive on Clean DICOMs while showing substantially stronger robustness in Video ZS and Video FT settings, which is the central focus of this work. We will avoid overstatement, clearly mark AUC as the primary metric due to class imbalance, report 95\% confidence intervals for key metrics, and perform statistical significance tests to examine whether observed differences are significant. 5.Realism, privacy, and clinical use (R2, R4): Our acquisition uses a controlled browsing agent with stochastic scrolling speed, occasional pauses, and limited reversals to generate screen-recorded videos with variable temporal patterns. We do not claim that this fully matches real clinician browsing; rather, it is a first controlled testbed for studying whether diagnostic information survives camera capture. We will clarify this limitation and note clinician-recorded videos, diverse devices/viewers, and broader acquisition settings as future validation. SVL reduces privacy risk by cropping peripheral UI/environmental regions, but screen-recorded medical data still require appropriate privacy protection and regulatory compliance.




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.

    All reviewers agreed that the idea behind the paper is sound and highly innovative (and I agree), yet that does not necessarily guarantee acceptance. The authors are encouraged to address the following issues raised by the reviewers and explain how they can be remedied in the frame of the rebuttal process. 1) Unfair comparison between ScreenCAD and baselines (R2, W1; R4, W2) has to be addressed –> this is highly critical 2) Results need more details in terms of display to be able to better assess their quality (R2, W2) 3) Agentic component is not evaluated and details are missing for agent scrolling screen (R4, W4) 4) Explanation on why CT only was evaluated (R1, W2; R4, W3)

    Note: Rx = reviewer x Wx = weakness x

  • 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.

    As previously mentioned, the paper is certainly innovative. Within the rebuttal, I could see that the most critical aspects (baseline comparison fairness) is not necessarily terrible. This aligns with reviewers 1 and 2.The remaining issues are also properly discussed and I really hope the authors are not lazy and really implement in the camera-ready version what they promised. Luckily for the authors, reviewer 4 did not justify the rejection such that based on my own evaluation and the one of the other two reviewers I give it a GO.



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 addresses an original and clinically relevant question: whether camera-captured scrolling videos of volumetric CT retain sufficient information for AI diagnosis when direct DICOM/PACS access is unavailable. Reviewers recognized the novelty, motivation, and potential clinical utility. The rebuttal substantially clarifies the main concern about baseline fairness by explaining that all methods used the same SVL-preprocessed video inputs and that CT-Net/CT-CLIP were also fine-tuned on captured videos. It also explains the role of MedDINOv3 versus the proposed task-token aggregation. Some limitations remain, especially the CT-only evaluation and controlled browsing-agent setup. Overall, based on the reviewer discussion, I recommend acceptance.



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.

    Two reviewers recommend acceptance after a successful rebuttal that addressed the primary concern centred around fairness of baseline methods.

    I agree that this paper formulates and investigates an interesting question (clinical value of video-recordings of high quality volumetric DICOM data). Even though investigating this question for other modalities than CT would be interesting, this paper is a solid first step with an innovative experimental setup.



back to top