Abstract

Artificial intelligence (AI) analysis of micro-ultrasound (μUS) has shown promise for prostate cancer (PCa) detection. However, most existing AI methods focus on the analysis of single μUS images in isolation. By contrast, expert μUS readers typically assess a full recorded video study, which provides three-dimensional context, to improve PCa detection compared to single-frame analysis. Inspired by this clinical workflow, we propose Compass, a novel AI methodology which models a μUS study as a stream of 2D images. Compass jointly integrates rotational sweep videos of the prostate with μUS frames acquired at the moment of biopsy, and performs evidence aggregation across the study using a transformer conditioned on the probe’s rotational angle. Finally, a decoder head predicts frame-level and study-level risk scores for the patient. The model is trained and evaluated using a multi-center clinical trial dataset of μUS studies, including continuous rotational scans of the prostate and videos captured during biopsy acquisition. We compare the proposed method to baseline AI methods from the literature and to risk scores provided by clinical experts. Our framework shows strong performance, highlighting the value of multi-view context for μUS PCa detection, and providing a potentially powerful tool to complement human expertise in μUS-based PCa diagnosis. Our code is available at: https://github.com/mharmanani/Compass.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5885_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/mharmanani/Compass

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WilPau_Compass_MICCAI2026,
        author = { Wilson, Paul F. R. AND Harmanani, Mohamed AND Guo, Zhuoxin AND Dzikunu, Obed K. AND Cash, Hannes AND Kinnaird, Adam AND Wodlinger, Brian AND Abolmaesumi, Purang AND Mousavi, Parvin},
        title = { { Compass: Prostate Cancer Detection Needs Multi-view Context } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper is a study-level framework that overcomes the limitation of existing AI methods that analyze individual 2D frames in isolation, by jointly leveraging rotational sweep videos and biopsy frames, mirroring how clinicians actually interpret µUS studies. Its key distinguishing feature is the injection of probe roll angle as positional encoding into the transformer. On a 118-patient multi-center dataset, it achieves a patient-level AUROC of 87.5% and Sen@60 of 92.9%, substantially outperforming existing baselines and expert PRI-MUS scores.

    The main contributions can be summarized as three: (1) reframing µUS prostate cancer detection as a multi-view study-level task rather than single-frame classification, (2) proposing a mixed-token transformer architecture that encodes probe roll angle via sinusoidal positional encoding to capture spatial relationships between sweep and biopsy tokens, and (3) validating on a multi-center clinical dataset with superior patient-level performance over both SOTA methods and human experts.

  • 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 identifies a meaningful gap between clinical practice (interpreting full 3D sweeps) and existing AI methods (single-frame analysis), and addresses it with a dual-branch architecture. The roll-angle sinusoidal PE is physically grounded, and the ablation study (Table 2) convincingly isolates each component’s contribution.

    2.Compass achieves a patient-level AUROC of 87.5% and Sen@60 of 92.9%, substantially outperforming the strongest baseline (ProstNFound+, 78.8%) and expert PRI-MUS (79.6%). The comparison spans frame-level, MIL, and video baselines, providing a thorough evaluation landscape.

    3.The observation that Compass excels at patient-level triage while PRI-MUS remains stronger at core-level characterization (Figure 2) is practically valuable, framing the method as complementary to expert judgment rather than a replacement.

    4.The authors provide an anonymous code repository, supporting reproducibility and enabling verification of the proposed method.

  • 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 5-fold CV results report only mean metrics without standard deviations or confidence intervals. With 118 patients, fold-level variance could be substantial, making it unclear whether the reported improvements are statistically significant.

    2.The 118-patient cohort from 2 centers is relatively small, and no external validation is provided. Moreover, the Conclusion does not discuss any limitations, including dataset size, the biopsy-level performance gap (71.0 vs 76.7), dependence on a frozen backbone, and the requirement for IMU roll angles. This is a notable omission for a clinical application paper.

    3.Sinusoidal PE, cross-attention, and [CLS] token-based transformer fusion are all well-established techniques. The paper should more clearly distinguish what is architecturally new from what is a thoughtful combination of existing methods applied to the µUS domain.

    1. Duplicated “multi-view multi-view framework,” “Mulit-view” typo in keywords, “Pca” instead of “PCa,” grammatical error in “we define to csPCa,” and undefined abbreviations (MIL, mpMRI).
  • 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 makes a well-motivated methodological and clinical contribution by reframing µUS PCa detection as a study-level multi-evidence reasoning task. The main concerns (lack of statistical validation and limitations discussion) are addressable in revision and do not undermine the core contribution. The novelty lies in the problem formulation and effective integration of multi-view context rather than individual components, which is appropriate for a clinical AI paper. Overall, this work should be of broad interest to the MICCAI community.

  • 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 rebuttal addresses my two main concerns. The authors committed to adding standard deviations across folds and an explicit limitations section discussing cohort size, the biopsy-level performance gap, and the need for prospective validation. They also clarified that the novelty lies in the problem formulation and the specific integration of sweeps, biopsy frames, and roll-angle geometry rather than individual components, which is consistent with my original assessment. The PRI-MUS comparison protocol (per-patient maximum, following OPTIMUM) and the framing of Compass as complementary to expert scoring are reasonable. My initial recommendation was Accept independent of rebuttal, and the rebuttal reinforces this.



Review #2

  • Please describe the contribution of the paper

    This paper makes a clinically meaningful shift from isolated-frame micro-ultrasound analysis to study-level reasoning for prostate cancer detection. It proposes Compass, a dual-branch framework that jointly models rotational prostate sweep videos and biopsy-time micro-ultrasound frames, conditions the fusion on probe roll angle, and predicts both patient-level and biopsy-level csPCa risk. The main contribution is showing that explicitly leveraging multi-view context from the full study can substantially improve patient-level detection compared with strong single-frame, MIL, and video-based baselines.

  • 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 addresses an important gap between how current AI systems often analyze isolated µUS frames and how clinicians actually interpret complete µUS studies with sweep context. (2) The use of sweep frames, biopsy-time frames, roll-angle encoding, and mixed-token transformer fusion is technically sensible for this acquisition setting. (3) The reported patient-level AUROC and sensitivity are strong, and the ablation study suggests that sweep information and cross-branch fusion both contribute meaningfully. (4) The paper is readable, the figures are helpful, and the clinical motivation is easy to follow.

  • 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 manuscript’s own related work already includes cine-loop modeling, temporal/spatio-temporal ultrasound modeling, heuristic slice-level aggregation, and strong µUS foundation-model approaches. Therefore, the main novelty here is the specific sweep+biopsy fusion design, rather than the broader idea of moving beyond isolated single-frame analysis. (2) The paper reports 5-fold patient-level cross-validation results, but only aggregate point estimates are shown. Given the small cohort (118 patients) and the two-center setting with different biopsy routes, the paper should report fold-wise mean±std or confidence intervals for all main metrics. Ideally, the authors should also provide per-center results or a leave-one-center-out analysis. (3) The manuscript states that results are reported at the epoch with the best mean patient AUROC across folds, but it is unclear how this epoch was selected and whether a separate validation split or nested cross-validation was used. The thresholding protocol for balanced accuracy and Sen@60 should also be described explicitly. As written, this raises concern about optimistic model selection. (4) Table 1 shows that Compass is strongest on patient-level AUROC and Sen@60, but it is not the best method on patient-level balanced accuracy, and it underperforms expert PRI-MUS at the biopsy/core level. The abstract, results, and conclusion should therefore avoid implying uniform superiority. Similarly, the conclusion mentions targeted biopsy planning, but since the model uses biopsy-time frames, the most immediate use case appears closer to intra-procedural decision support than true pre-biopsy planning. (5) Compass appears to use richer inputs than several baselines (sweeps, biopsy frames, roll-angle information, and metadata), so the paper should clearly specify which inputs and metadata each baseline receives, whether all methods share the same splits/training budget/model-selection protocol, and how the patient-level PRI-MUS comparator is derived. As written, some of the comparisons do not appear fully apples-to-apples. In fact, the no-sweeps ablation may be the fairest internal comparison and should be emphasized more strongly. (6) Some methodological descriptions are inconsistent. The text says the biopsy prompts use Age and PSA, Figure 1 also shows PSA density, and the dataset description additionally mentions family history. The paper should provide a precise and consistent description of which clinical variables are used in each experiment.

  • 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?

    I am currently slightly below the acceptance threshold. The paper studies an important clinical problem, and the idea of incorporating sweep-level multi-view context is interesting and promising. The patient-level AUROC and sensitivity gains are notable, and the ablations support the claim that sweep information and cross-branch fusion matter. However, I do not yet have enough confidence in the current version for acceptance. The paper would need more rigorous and transparent reporting of uncertainty (e.g., five-fold mean±std / confidence intervals), a clearer validation and model-selection protocol, more accurate claim language, and a more explicit discussion of comparison fairness. In particular, because Compass uses richer inputs than several baselines and does not dominate on all reported metrics, the paper should be more careful in how it frames its advantages. A rebuttal that directly clarifies these issues would improve my confidence, but based on the current draft I lean toward weak reject.

  • 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 have clarified that the main contribution is not the broad idea of moving beyond isolated single-frame analysis, but the specific formulation of jointly modeling rotational µUS sweeps, biopsy-time frames, and roll-angle geometry within a single study-level framework. This is a reasonable and meaningful contribution in the µUS prostate cancer detection setting. The reported patient-level performance is strong, especially for AUROC and Sen@60, and the no-sweeps ablation provides useful evidence that sweep information contributes complementary signal beyond biopsy-frame-only modeling. The rebuttal also addresses the issue of overclaiming. The authors acknowledge that Compass does not uniformly outperform all comparators, especially at the core level against expert PRI-MUS, and they commit to tempering the manuscript by positioning Compass as a complementary tool rather than a replacement for expert judgment. I find this clarification acceptable. Similarly, while Compass uses richer inputs than several baselines, the authors reasonably argue that leveraging these additional inputs is part of the proposed contribution. The final paper should still clearly specify which inputs each baseline receives and avoid implying a fully apples-to-apples comparison where the input settings differ. Some weaknesses remain, including the small cohort size, the need for clearer uncertainty reporting, a more explicit description of model selection and thresholding, and consistent reporting of the clinical variables used. However, these issues mainly concern transparency, presentation, and claim calibration, and can be addressed in the camera-ready version without requiring new experiments. Overall, the paper presents a clinically motivated and technically coherent framework with promising patient-level results and supportive ablation evidence. I therefore recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    The main contribution of the paper is the introduction of Compass, a novel AI framework for prostate cancer detection from micro-ultrasound (µUS) that moves beyond traditional single-image analysis to multi-view, study-level reasoning.

  • 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 tackles a clinically meaningful and well-motivated problem, aligning closely with real diagnostic workflows. Rather than treating µUS as isolated 2D frames, it models a study as a continuous stream of multi-view observations, better reflecting how clinicians interpret prostate sweeps.

    A second strength is the integration of complementary data sources, including rotational sweep frames, biopsy frames, and clinical metadata.

    The method also incorporates acquisition geometry through rotation-aware encoding, preserving spatial structure across frames.

    Empirically, the paper demonstrates strong and consistent performance gains, including improvements in patient-level AUROC and sensitivity over a wide range of baselines (frame-based, MIL, video models, and expert scoring).

  • 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.Limited novelty in core modeling components. While the paper positions Compass as a novel framework, many of its building blocks are incremental combinations of existing ideas. The use of transformers for aggregating multiple instances (frames, slices, or views) is already well explored in medical imaging. The use of foundation model backbones (ProstNFound+) with prompt conditioning is directly inherited from prior work.

    2.The method is evaluated on only 118 patients from two centers , which is relatively small. No external validation set beyond the two centers are provided. This is particularly important since prior work (e.g., ProstNFound+, Cinepro) emphasizes robust training and domain generalization.

    3.There are also limited demonstrations of clinical feasibility and workflow integration. Although the paper motivates alignment with clinical workflow, actual clinical usability is not rigorously evaluated. There is no study on reader-AI collaboration, or time efficiency or workflow integration.

    4.Compass explicitly incorporates biopsy frames as input. This raises concerns for applicability for pre-biopsy decision-making and potential information leakage, since biopsy targets are already clinician-selected suspicious regions

    5.There is no failure case analysis. It would be insightful to know performance across prostate sizes, lesion locations or patient subgroups.

  • 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?

    The paper presents a well-motivated and empirically strong system, but the contribution is primarily incremental, and the experimental validation is not sufficiently comprehensive to support strong claims of clinical impact or general applicability.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    After considering the authors’ rebuttal, I maintain my initial recommendation of reject. I appreciate the clarifications, but the rebuttal does not sufficiently address my main concerns regarding novelty, validation robustness, and clinical workflow applicability.

    First, although the authors emphasize that their framework uniquely combines rotational sweeps, biopsy frames, and roll-angle geometry, the core modeling components remain largely incremental combinations of established techniques. Second, while the two-center 5-fold cross-validation is reasonable for an initial study, the small cohort size and absence of an independent external validation set limit the strength of the paper’s generalizability and clinical-impact claims. Third, the rebuttal’s proposed pre-biopsy deployment strategy is not the setting actually evaluated in the manuscript. The current model still uses clinician-selected biopsy frames.

    Overall, the paper is well motivated and shows promising initial results, but the technical contribution is incremental and important validation questions remain insufficiently addressed. Therefore, my assessment remains unchanged.



Author Feedback

We thank all reviewers for their constructive feedback. Reviewers have recognized our strong clinical motivation (R1, R2, R3), “strong and consistent performance gains” over a “thorough evaluation landscape” (R1, R2, R3); and the “practical value” of Compass (R1, R2) as a “complementary tool to expert judgment” (R1). We address major concerns below.

NOVELTY (R2, R3): No prior µUS work jointly models rotational sweeps, biopsy frames, and roll-angle geometry, acquired during a biopsy study, in a single framework. Therefore, our problem formulation, model design, and empirical findings (that study-level reasoning substantially outperforms every available baseline), all constitute novel contributions. While sinusoidal encodings, cross-attention, and transformer fusion are used in other literature, their combination in this specific formulation and problem setting is novel. We will clarify this in revisions.

STATISTICAL UNCERTAINTY (R1, R2, R3): We agree that uncertainty reporting would improve the paper. We have this information and will amend the tables to include the std. dev. of the reported metrics.

METRICS & MODEL SELECTION (R2): The same selection protocol was applied to all methods: we selected the epoch with the highest mean validation AUROC across folds. Sen@60 is the sensitivity at the first operating point where specificity exceeds 60%. Balanced Acc. is derived from a single point estimate of Sens. and Spec., and is sensitive to the choice of threshold. At deployment, the threshold would be calibrated to the clinical setting. AUROC is more robust as it summarizes performance across thresholds. We will clarify these points in revisions.

COMPARISON FAIRNESS (R2): We clarify that integrating richer inputs (sweeps and roll angles alongside still frames) is part of our contribution. These inputs are easy to obtain during a µUS study, but no prior work proposes a framework to leverage them effectively. This gap motivated Compass, whose multi-stage architecture (i) exploits these rich inputs for the first time, (ii) exploits them better than baseline alternatives at each stage (frame-level, MIL, video), and (iii) is strongly supported by ablations. For the PRI-MUS comparison, we use the maximum PRI-MUS per patient, consistent with the landmark OPTIMUM clinical trial on µUS. The comparisons are thus fair: each baseline receives the inputs it was designed for.

CORE-LEVEL PERFORMANCE (R1, R2): The reviewers observe that Compass does not outperform PRI-MUS on cores. We emphasize that Compass is not intended to replace PRI-MUS, but complement it. Matching the study’s PRI-MUS raters (experts with years of experience) would be remarkable for an AI system. Even when PRI-MUS experts are available, our system complements them with stronger patient-level performance. Further, PRI-MUS scoring relies on specialized expertise unavailable in many clinical settings outside large urban centers; in these cases, Compass may provide a viable alternative. We will amend the manuscript to make these explicit and temper claims.

CLINICAL FEASIBILITY (R2, R3): We only require labelled biopsy-time frames to train and test the model. At deployment, frames can be obtained by capturing images from typical biopsy locations before the procedure, so the model can be used in routine clinical workflows and pre-biopsy screening.

DATASET SIZE (R1, R2, R3): We are aware that cohort size is a limitation. Our use of five-fold cross-validation and data from two geographically distant centers improves the strength of our results within the constraints of dataset size. We also note that 118 patients is consistent with other published µUS AI studies used for model development (e.g. MicroSegNet, ProMUS-NET). The cohort is therefore sufficient to demonstrate feasibility, but we will state the limitation and need for prospective validation in future work in revisions.

We hope our clarifications address your concerns, and will incorporate them in the final revision.




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.

    Given the mixed reviews, I recommend inviting a rebuttal. The authors should clarify several key issues: statistical uncertainty of the 5-fold results, validation/model-selection protocol, fairness of comparisons given richer Compass inputs, and whether the use of biopsy-time frames limits pre-biopsy applicability or introduces bias. The rebuttal should also address limited cohort size, lack of external validation, core-level underperformance relative to PRI-MUS, and provide clearer limitations and failure-case analysis.

  • 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.

    Following the rebuttal, two reviewers recommended acceptance, noting that the authors clarified the novelty as the specific integration of sweep, biopsy, and roll-angle information, agreed to report fold-level variability, clarified model selection and thresholding, and appropriately reframed Compass as complementary to expert PRI-MUS rather than uniformly superior. However, one reviewer still had concerns about the core components remaining incremental, the cohort being small, no independent external validation. Overall, while validation limitations remain, the rebuttal addressed key concerns and indicate a promising clinically motivated contribution. I therefore recommend acceptance.



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.

    Two reviewers are happy with the rebuttal. Although some limitations, such as a small test cohort and model selection justification, still exist, two reviewers agree that this is a good system paper. So I would 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.

    The authors addressed main concerns provided by two reviewers. I would recommend to include it in camera-ready.



back to top