Abstract

Clinical prostate multi-parametric MRI relies heavily on high-quality diffusion-weighted imaging (DWI), yet reading DWI is frequently compromised by geometric distortion, often caused by rectal air. Assessing quality via the PI-QUAL scoring system is an emerging clinical standard, but it is subjective, time-consuming and suffers from a class imbalance where low-quality cases are diverse and relatively scarce. Using the PRIME clinical trial as an example, there are 6% images with PI-QUAL scores lower than 4, 87% of DWI issues are due to distortion. Many of the other clinical quality issues are under-represented. To address this common dual-scarcity of annotated clinical data, we propose a few-shot biparametric prototypical network for automated image quality assessment (IQA). Our framework utilizes a dual-branch 3D ResNet to fuse T2-weighted and DWI features, providing anatomical context to distinguish true morphology from distortion. To handle real-world heterogeneity, we introduce feature-wise linear modulation (FiLM) and a gradient reversal layer (GRL) to align feature distributions conditioned on varying b-values while suppressing acquisition-related biases. We demonstrate that a model meta-trained solely on comparatively objective, readily obtainable distortion labels can effectively adapt to predicting complex, multi-factorial clinical quality scores such as PI-QUAL using only five representative samples. Experimental results on two datasets show that our method significantly outperforms few-shot learning baselines for this challenging IQA task, offering a practically feasible and data-efficient solution for standardizing prostate MRI quality control in clinical workflows. Code is available at https://github.com/yucheng722/Proto-FM-IQA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3982_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yucheng722/Proto-FM-IQA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{TanYuc_Bridging_MICCAI2026,
        author = { Tang, Yucheng AND Ng, Alexander AND Yan, Wen AND Thorley, Natasha AND Rajwa, Pawel AND Wang, Yipei AND Asif, Aqua AND Allen, Clare AND Dickinson, Louise AND Giganti, Francesco AND Punwani, Shonit AND Alexander, Daniel C. AND Kasivisvanathan, Veeru AND Hu, Yipeng},
        title = { { Bridging Single Distortion Artifacts and Multifactorial Clinical Quality: Few-Shot Biparametric MRI Quality Assessment via Distortion-Trained Prototypical Networks } },
        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 main contribution of this paper lies in bridging few-shot learning to a realistic and clinically relevant MRI image quality assessment (IQA) problem under small data scenario, which is characterized by severe class imbalance and heterogeneous quality factors.

  • 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 a clinically important and well-motivated problem of prostate MRI quality assessment under the PI-QUAL standard, with a focus on real-world challenges such as class imbalance and the scarcity of diverse low-quality samples.

    2.The two-stage learning strategy is practical, leveraging comparatively objective distortion labels to facilitate learning for a more complex and clinically relevant task, with potential to generalize to other medical imaging scenarios.

  • 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.While the paper is motivated by data scarcity, the experimental dataset (Dataset 1) contains 1027 samples, which does not represent a strictly low-data regime. The results in Table 1 show that a standard fully supervised baseline (FSL), trained on the full training split, already achieves strong performance and outperforms the proposed method. This raises the concern that the few-shot formulation may not provide practical benefits in this scenario and appears somewhat driven by methodology rather than data constraints.

    2.A more realistic evaluation protocol is needed, where instead of constructing episodic tasks on the test set, the meta-trained model should also be evaluated under a direct inference setting by being applied to the test data without forming support/query splits. This setting is more consistent with real-world deployment and aligns with recent trends in few-shot learning, where non-episodic evaluation has become increasingly common.

    3.It is important to report the performance of a fully supervised model on the PRIME dataset (FSL baseline). Regardless of whether the results are strong or not, including this baseline would provide a clearer comparison of the benefits of the proposed approach.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    My recommendation is mainly based on concerns about the necessity of the few-shot formulation and the evaluation design. Since the task is claimed to be clinically relevant, a standard fully supervised baseline already achieves strong performance and even outperforms the proposed method on Dataset 1 (the baseline result on PRIME is not reported), which raises concerns about the benefit of introducing a meta-learning formulation, as it appears to focus more on methodological validation than on clinical utility.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    Thank you for the rebuttal. My main concerns have been addressed, and I raise my score to 4.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a two-stage learning strategy for automated prostate MRI quality control, transferring distortion-aware representations to clinically specific assessment through meta-training and few-shot adaptation.

  • 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 develops a few-shot framework tailored for prostate MRI quality assessment, effectively addressing the scarcity of clinically annotated data. 2.It leverages structural T2WI as an anatomical reference to help distinguish true anatomical variation from geometric distortion in DWI. 3.It introduces feature alignment modules (FiLM and GRL) to improve robustness to variations in scanning parameters.

  • 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 paper is presented as prostate MRI quality assessment, but the method is mainly built around distortion representation learning and is clinically validated only on binary PI-QUAL classification. 2.The method is sensitive to support-set selection, which may affect robustness in few-shot adaptation.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper presents a practical and reasonably validated method for an important clinical problem. Still, its empirical scope is somewhat limited, and the paper would be strengthened by better aligning its claims with the current validation scope.

  • 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 partly address the support-set sensitivity concern by explaining that support samples are clinically curated. However, a gap remains between the broad claim of general clinical MRI quality assessment and the actual validation, which is still relatively limited. Overall, I lean toward Weak Accept.



Review #3

  • Please describe the contribution of the paper

    This paper studies automated quality assessment for biparametric prostate MRI under limited annotation. The main idea is to first meta-train a few-shot prototypical network on a distortion detection task using DWI and T2WI, and then transfer the learned representation to a more clinically oriented quality assessment task, namely PI-QUAL prediction, using only a small number of support samples. The method combines a dual-branch 3D ResNet with FiLM-based conditioning and GRL-based alignment to improve robustness across acquisition settings and datasets.

  • 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 a clinically relevant problem. MRI quality assessment, especially for DWI, is important in prostate imaging and has clear practical value. 2.The overall motivation is sensible. Using a more objective and easier-to-label distortion task as a source task for transfer to a more subjective and data-scarce clinical quality task is an interesting idea.

  • 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 biggest weakness is that the paper overclaims what it actually demonstrates. The proposed method is trained entirely on distortion labels, yet the paper frames the contribution as addressing broader clinical quality assessment and PI-QUAL prediction. However, PI-QUAL is a multifactorial concept, not simply a distortion label. Since the PRIME cohort is also heavily dominated by distortion-related low-quality cases, the current results mainly show that distortion-sensitive features transfer to another distortion-heavy setting. That is much narrower than the paper’s broader clinical claim. 2.The evidence for generalization to real PI-QUAL assessment is therefore limited. The paper does not show convincing analysis of non-distortion low-quality cases, so it is unclear whether the model is actually learning broader image quality or mostly detecting distortion. 3.The few-shot adaptation setting also limits practical value. The method assumes five support samples per class for adaptation, including rare low-quality cases. In realistic deployment, this requirement may still be nontrivial, especially in well-controlled clinical cohorts where low-quality scans are uncommon. The paper does not discuss this limitation in enough depth. 4.The results may be quite sensitive to support set selection, but this is not studied carefully. In few-shot learning, performance can vary substantially depending on which examples are chosen as support samples, and the current paper does not provide enough evidence that the reported gains are robust to that choice. External validation is present, but still limited. The number of low-quality and distorted cases in PRIME is small, which makes it difficult to judge how stable the conclusions would be in more diverse populations and acquisition settings.

    Overall, while the idea is interesting, the current empirical evidence is not strong enough to support the paper’s broader claims about clinical quality assessment.

  • 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.

    (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 scored the paper below the acceptance threshold mainly because I do not think the current evidence fully supports the paper’s central claims. The method is trained on distortion labels, but the paper presents the contribution in terms of broader clinical quality assessment and PI-QUAL prediction. Given that the downstream evaluation cohort is also largely dominated by distortion-related low-quality cases, the results mainly demonstrate transfer of distortion-sensitive features rather than general image quality assessment. This makes the main claim feel overstated. In addition, the practical deployment implications of the five-shot adaptation setup are not fully addressed, and the sensitivity to support set selection is not analyzed in enough detail. Overall, the paper is interesting and reasonably well executed, but I find the scope of the claims somewhat ahead of the actual evidence, which leads me to a weak reject recommendation.

  • 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.

    I appreciate the authors’ response. The clarification on the motivation for few-shot adaptation is helpful, especially the distinction between source-domain FSL performance and the intended target-domain/cross-task adaptation setting. However, my main concerns are only partially addressed.

    First, the response still does not provide a systematic analysis of support set sensitivity. The authors clarify that the five support samples are curated to cover b-value ranges and different low-quality patterns, but this also suggests that the method depends on expert-selected representative samples. Since the paper itself acknowledges sensitivity to support set selection, I still think this should be evaluated more carefully rather than only described as part of the clinical workflow.

    Second, the overclaiming concern remains. The authors clarify that some low PI-QUAL cases involve T2 quality issues and that distortion is a major determinant of DWI-related quality degradation. However, the current evidence still mainly demonstrates transfer from distortion-label training to a distortion-dominated PI-QUAL setting. The paper does not provide a convincing subgroup analysis of non-distortion low-quality cases, so it remains unclear whether the model learns broader multifactorial PI-QUAL quality or primarily detects distortion-related degradation.

    Third, the statement that the method uses only distortion-labelled training data should be phrased carefully. While the meta-training labels are distortion labels, the PI-QUAL adaptation stage still requires five PI-QUAL-labelled support samples per class. This is a reasonable few-shot setting, but it is not label-free PI-QUAL prediction.

    Overall, the rebuttal improves the explanation of the intended setting, but it does not fully resolve my concerns about support-set robustness and the strength of the claim that the method performs general clinical PI-QUAL assessment. My assessment would remain largely unchanged unless these claims are toned down or additional analyses are added.



Author Feedback

We sincerely thank the Area Chair (AC) and Reviewers for their constructive feedback.

1.Necessity of Few-Shot Adaptation (R1)

R1 argued that fully supervised learning (FSL) outperforms few-shot learning on Dataset 1 and therefore questioned the necessity of our method. Indeed, Dataset 1 is a source-domain meta-training cohort, where strong in-domain FSL performance is expected. We view this as an “oracle” upper bound requiring extensive target-domain annotations, rather than a practical solution in our intended data-constrained challenges, as in many real-world clinical applications. These require adaptation to new datasets and tasks (e.g. PRIME PI-QUAL assessment), where only a few low-quality (LQ) examples are available. In such settings, FSL becomes impractical and prone to severe overfitting, consistent with the poor FSL-5 performance on PRIME shown in Table 1.Moreover, directly applying the Dataset-1-trained FSL model to PRIME suffers from substantial domain shift caused by scanner and protocol variability across centers. This motivates few-shot learning that can transfer source-domain knowledge under clinically and practically limited target-domain supervision.

2.Expected Realistic Clinical Setting of Episodic Protocols (R1,R2,R4)

We agree that, although it is still non-trivial as pointed out by the reviewer, our approach proposes a practical deployment scenario: adaptation is performed once for a new site/scanner/protocol using only five archived support scans. Non-episodic inference (R1 suggested) is certainly advantageous, but in current study, its trade-off to performance is generally poor and exemplified in Table 1.

Sensitivity to a very small number (here, five) support samples (R2 and R4) is indeed the challenge that many few-shot algorithms address, also in part motivated our proposed prototypical methods designed for this challenge. Practically in our clinical workflow, the support sets are curated (instead of random sampled) to cover a b-value range of 1400–2000 and different types of LQ images (e.g. distortion vs non-distortion). However, this step was performed independently prior to experiment and as demonstrated in our experiments, yielded promising IQA performance. Related comment from R4 noted that the limited number of LQ samples in PRIME may restrict conclusions regarding stability. We emphasize that PRIME is a large-scale, highly heterogeneous multi-center dataset (483 cases across scanners/protocols/institutions). The prevalence of LQ examples therefore reflects the real clinical scenario motivating our framework. Technically, the proposed approach leverages source-domain meta-learned representations while requiring only minimal target-domain calibration, improving robustness under scarce supervision.

3.”Overclaiming” (R4)

R4 noted that our model is trained on distortion labels but overclaims to solve general quality assessment. We first clarify that, in the PRIME trial, 29% of low PI-QUAL cases were associated with T2 quality issues and 87% among DWI-related quality issues were attributed to geometric distortion - a major determinant of clinical image quality in this application.

More importantly, we clarify that our target is clinical, diagnosis-relevant prostate MRI quality assessment, as measured (and thus validated) by PI-QUAL, on biparametric T2 and diffusion MR from real clinical cases. In fact, the purpose of our study is to assess such a general IQA (defined for this application), but using training data only being “distortion-labelled” - by domain adaptation - because the distortion labels are much feasibly and reliably to obtain for a larger training set as demonstrated in this study.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviewers raise points about an extant baseline with better performance, and issues with how realistic of a clinical setting the authors’ experiment actually is.

    I recommend that a rebuttal address whether these concerns are correct or not.

  • 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.

    Reviewer majority, R3 concerns seem overly critical, though valid. I find this paper to be borderline: a good method perhaps, but with clear weaknesses that remain unaddressed.



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 a clinically relevant and practically important problem: prostate MRI quality assessment under limited annotated data. The reviewers generally agreed that the work is technically sound, well motivated, and reasonably validated, with strengths including the distortion-aware transfer strategy, adaptation to low-data settings, and the integration of FiLM/GRL modules for robustness across acquisition settings. The study also provides external validation on a heterogeneous multi-center dataset and tackles a realistic challenge in prostate MRI quality control.

    The main concerns raised by reviewers focused on the scope of the claims rather than fundamental methodological flaws. In particular, reviewers questioned whether the current evidence fully supports broader claims of general PI-QUAL assessment beyond distortion-related quality degradation, and noted limited analysis of support-set sensitivity and deployment practicality. These concerns are valid and should be addressed through more careful positioning and discussion of limitations. However, the rebuttal clarified the intended use case more convincingly: the method is designed for cross-domain adaptation under severe annotation scarcity, rather than outperforming fully supervised learning in large in-domain settings. The authors also clarified the practical clinical workflow for support-set curation and the motivation for leveraging distortion labels as a scalable surrogate supervision signal.

    Importantly, none of the reviewers identified technical errors, invalid experiments, or major reproducibility concerns. After rebuttal, two reviewers moved to or maintained acceptance recommendations, while the remaining concerns primarily relate to claim calibration and the limited breadth of validation rather than a lack of scientific merit. Given the overall technical quality, clinical relevance, and positive reviewer consensus after rebuttal, 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.

    The reviewers had split opinions, reflecting the complexity of the task. I recommend acceptance of the study based on the overall quality and reviewers comments, but I strongly encourage the authors to consider the critical points raised by the reviewers.



back to top