Abstract

PET/CT imaging is evolving from a single-tracer modality to a multi-tracer paradigm, enabling precise characterization of diverse disease patterns. However, a core diagnostic challenge remains, i.e., distinguishing pathological lesions from physiological uptake, as both appear as high-intensity “hot spots.” This distinction is highly dependent on the anatomical location and tracer type, making automated segmentation across tracers particularly difficult. Standard Mixture-of-Experts (MoE) architectures theoretically support multi-task learning but typically rely on feature-driven gating. This leads to uncontrollable routing, because of failing to leverage the anatomical and tracer-specific priors essential for radiological interpretation. In this paper, we propose ProMoE, a Prototype-Driven Mixture-of-Experts framework for context-informed PET lesion segmentation across tracers. Unlike conventional MoE, ProMoE introduces a set of learnable expert prototypes that serve as anchors within a semantic embedding space to represent particular physiological contexts. During inference, each input is first embedded into this space according to anatomical and tracer-specific priors. Then, expert activation is performed driven by the similarity between the embedding and expert prototypes. This mechanism supports physiologically informed expert specialization and tracer-adaptive routing. Validated on a large-scale multi-center dataset of 1,366 PET/CT scans across 3 tracers 18F-FDG, 18F/68Ga-PSMA, and 18F-FAPI, ProMoE significantly outperforms the best state-of-the-art method by 5.97% in average DSC. Furthermore, it yields an 8.82% DSC improvement in prior-guided zero-shot transfer on 30 cases of an unseen tracer (CD70), underscoring its potential as a scalable, biologically grounded framework for prior-guided adaptation to new PET tracers. The code is publicly available at https://github.com/CMYXing/ProMoE.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0325_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CMYXing/ProMoE

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ShuYir_TracerAdaptive_MICCAI2026,
        author = { Shu, Yiran AND Liu, Jingyu AND Yu, Mingyang AND Wang, Xinlu AND Li, Guiyu AND Wei, Weijun AND Li, Lianghua AND Chen, Yanbo AND Gao, Yaozong AND Shen, Dinggang},
        title = { { Tracer-Adaptive Expert Learning for Generalizable PET Lesion Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    ProMoE addresses the challenge of distinguishing pathological lesions from physiological uptake across different PET tracers. It introduces a Prototype-Driven Mixture-of-Experts (MoE) where routing is guided by an explicit semantic embedding space based on anatomical location and expected physiological uptake.

  • 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.Moving away from standard feature-driven MoE gating to a biologically grounded, prototype-driven routing mechanism is relatively innovative. 2.The evaluation on 1,366 scans across three diverse tracers is comprehensive. The zero-shot generalization testing on a completely unseen tracer (CD70) proves the model’s scalability.

  • 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.Cascading Error Vulnerability: The routing logic strictly depends on anatomical occupancy maps generated by a frozen CT backbone pre-trained on TotalSegmentator. If the CT segmentation fails due to poor contrast, metal artifacts, or unusual patient positioning, the physiological expectation embedding will be flawed, causing the MoE routing to collapse. The paper lacks an ablation study analyzing the framework’s robustness to anatomical segmentation errors. 2.Scalability Bottleneck: The ablation study reveals that increasing the number of experts to K=36 degrades overall performance. The authors hypothesize that excessive sparsity hinders shared representation learning. For a foundation-level architecture aiming to handle an expanding array of tracers, this capacity limit must be addressed or mathematically bounded.

  • 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 does not provide sufficient information for 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.

    (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?

    ProMoE introduces a highly innovative conceptual shift by moving away from standard, black-box feature-driven MoE gating toward a biologically grounded, prototype-driven routing mechanism. This is a clever and interpretable use of the MoE architecture. The zero-shot generalization results on an entirely unseen tracer (CD70) prove the model’s scalability. The limitations are that the architecture relies heavily on a frozen CT backbone for anatomical occupancy mapping; without an ablation study addressing how the MoE routing behaves when this upstream CT segmentation fails, the pipeline has a unaddressed vulnerability.

  • 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.

    My decision remains the same



Review #2

  • Please describe the contribution of the paper

    This paper presents ProMoE, a prototype-driven Mixture-of-Experts framework for multi-tracer PET lesion segmentation. The core idea is to replace standard feature-driven MoE routing with routing guided by anatomical context from CT and tracer-specific physiological uptake priors. Concretely, the method constructs a Physiological Expectation Embedding (PE2) from organ occupancy maps and LLM-curated tracer-specific uptake levels, matches this embedding to learnable expert prototypes via a joint anatomical-physiological similarity score, and uses the resulting routing weights to modulate PET features through a dual-stage anatomy-aware (AaFM) and physiology-aware (PaFM) design. The paper reports strong results on a relatively large multi-tracer PET/CT dataset and demonstrates promising performance on an unseen tracer (CD70).

  • 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 technically challenging problem. Multi-tracer PET lesion segmentation is genuinely difficult because lesions and physiological uptake often share similar high-intensity appearances, and correct interpretation inherently depends on both anatomical location and tracer type. The problem formulation is well-scoped and the clinical motivation is clearly articulated throughout. (2) The core motivation is well-grounded and matches clinical reasoning. Guiding expert routing with anatomical context and tracer-specific physiological priors is a more principled approach than standard feature-only MoE gating, and the design mirrors how nuclear medicine physicians interpret PET studies in practice. The “conflicting gradient” problem identified in conventional MoE under multi-tracer settings is a genuine and underappreciated failure mode that this work directly addresses. (3) The method has meaningful novelty at the application level. The combination of PE2 embedding, prototype-driven routing with a joint anatomical-physiological similarity metric, and dual-stage feature modulation is a thoughtful and coherent formulation. The use of an LLM to curate a multi-tracer physiological atlas as structured prior knowledge is an interesting design choice, and the K-Medoids initialization strategy for expert prototypes reflects careful attention to training stability. (4) The experimental results are compelling, particularly in terms of parameter efficiency. ProMoE achieves state-of-the-art performance with only 6M parameters, approximately 1.5% of nnU-Net, while outperforming all multi-tracer baselines by a substantial margin, especially in the PSMA domain (+8.93% DSC over SwinUNETR). The result that ProMoE surpasses single-tracer specialists in the data-scarce FAPI setting further supports the value of the proposed cross-tracer knowledge sharing mechanism.

  • 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 “zero-shot generalization” claim requires more careful qualification. The paper claims zero-shot generalization to the unseen CD70 tracer, which is partially accurate in that no CD70 training data is used. However, the method relies on tracer-specific physiological priors for CD70 at test time, encoded in the PE2 embedding. This makes the setting more precisely described as prior-guided zero-shot rather than strictly blind zero-shot generalization. The authors should clarify how the CD70 physiological prior was constructed and whether any CD70 uptake characteristics were known a priori, as this directly affects the validity of the zero-shot claim. (2) The contribution of the routing mechanism itself is not fully isolated from the backbone advantage. ProMoE uses a frozen CT backbone pre-trained on TotalSegmentator, providing structured anatomical supervision not available to the baselines. While all methods receive PET+CT as input, the quality of anatomical conditioning differs substantially, making it difficult to attribute performance gains specifically to the proposed prototype-driven routing. A controlled baseline sharing the same frozen CT backbone but using standard feature-driven routing would more cleanly isolate the routing contribution. The ablation in Table 2 partially addresses this within the ProMoE family, but does not resolve the fairness concern against external baselines. (3) Several technical details are insufficiently specified. The binarization threshold in 1(v), the motivation for squaring S_Phy in the joint similarity score, the selection criterion for the hard threshold theta, and whether K=12 was chosen via cross-validation are all left unspecified. The squaring of S_Phy in particular is a non-trivial design decision that warrants either theoretical justification or an ablation. These gaps modestly limit reproducibility and confidence in the formulation. (4) The generalization claims are somewhat stronger than the evidence supports. The unseen-tracer test set contains only 30 CD70 cases, which is a small sample for drawing strong conclusions about zero-shot generalization. Center-disjoint evaluation is also absent, making it difficult to fully assess cross-domain robustness. The generalization claims in the abstract and conclusion should be toned down to better reflect the scope of the current evidence.

  • 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 provide sufficient information for 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.

    (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?

    This paper tackles an important problem and proposes a clinically meaningful idea: guiding MoE routing with anatomical and tracer-specific physiological priors instead of relying only on image features. I think this is a sensible direction for multi-tracer PET lesion segmentation, and the reported results are promising. The method also appears to work well in the unified one-for-all setting, which is a practical strength. That said, I am still slightly below the acceptance threshold. The main issue is that the experimental claims are somewhat stronger than what is fully supported by the current evaluation. In particular, the “zero-shot” setting seems to rely on tracer-specific prior knowledge at test time, which makes the claim less clean than presented. Also, the proposed method uses additional structured priors and a frozen organ-aware CT branch, while the baselines do not appear to have equivalent conditioning, so the comparison is not fully controlled. Finally, several implementation details are unclear, which limits reproducibility and makes some design choices harder to assess. Overall, I think the paper has a good idea and clear potential, but the current version still needs stronger experimental clarification and more precise presentation to justify acceptance.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    After reading the rebuttal, I am willing to change my recommendation to Accept. The authors addressed my main concerns reasonably well, especially by agreeing to qualify the “zero-shot” claim as “prior-guided zero-shot transfer,” clarifying the CD70 setting and atlas construction, and providing missing implementation details such as the threshold choice, K selection, and the use of the squared physiological similarity term. The rebuttal also partially alleviates my concern about comparison fairness by clarifying that the FiLM w/o MoE ablation already uses the frozen CT backbone, and that the proposed prototype-driven routing outperforms MLP-MoE under the same expert setting.



Review #3

  • Please describe the contribution of the paper

    The authors propose a multi-tracer PET lesion segmentation approach that reframes MoE routing as an anatomy- and physiology-aware matching problem, rather than learning physiology implicitly through data-driven clustering. The proposed method incorporates a structured inductive bias that improves performance while using fewer parameters. The approach is validated against strong baselines, supported by ablation studies, and demonstrates promising zero-shot transfer to an unseen tracer, highlighting its potential for clinical generalisation.

  • 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.

    Atlas-based encoding of anatomical uptake expectations enables context-aware interpretation, and the method shows strong robustness in physiologically confounded regimes (e.g. PSMA, FAPI), addressing a key limitation of existing approaches.

  • 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 paper’s core claim of improved physiological–pathological disambiguation in multi-tracer setup necessitates evaluation of organ-wise false positive rates and region-specific error breakdowns. The current reliance on aggregate DSC and NSD, while standard, leaves the link between the proposed mechanism and the reported improvements largely implicit, particularly for FDG, where gains are marginal. The qualitative examples in Fig. 3 are illustrative but limited in scope, and do not fully substantiate the claimed behaviour across the dataset.

    The ablation in Table 2 suggests that the MoE routing contributes unevenly across tracers. For FDG, the FiLM-conditioned baseline (K=1) performs within ~2% DSC/NSD of the full ProMoE, indicating that atlas-based conditioning alone nearly saturates performance for metabolically diffuse tracers. The more substantial gains on PSMA and FAPI are encouraging, but it would strengthen the paper to clarify whether these stem from the prototype-driven routing specifically or from the increased model capacity that additional experts provide. An experiment comparing a single-branch architecture with tracer-conditional FiLM parameters against the full ProMoE at matched parameter count would help disentangle these factors.

    Reliability of generated uptake priors, and uncertainty from LLM based physiological uptake atlas impacts the reproducibility and robustness, and would benefit from validation.

  • 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 provide sufficient information for 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

    I would encourage the authors to justify why MoE routing is necessary beyond what the physiological prior injection already provides. The current results suggest this distinction is tracer-dependent (cd70 is similar in the confounding organ localised physiological uptake to PSMA and FAPI) and may not generalise as broadly as claimed.

  • 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 anatomical–physiological prior is a compelling aspect of the work, but the claim of improved physiological–pathological disambiguation is not supported by quantitative analysis, and the necessity of the MoE formulation is not convincingly established. I would welcome clarification on these points in the rebuttal.

  • 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 acknowledges the CD70 transferability assumption and recalibrates the zero-shot framing. The MoE necessity remains incompletely justified, but the FDG-dominance argument is mechanistically plausible.

    I would recommend amending the claim of ‘physiological-context-aware segmentation’ to ‘physiological-context-informed segmentation,’ as the reported evaluations are not sufficient to demonstrate awareness of physiology in the outputs.



Author Feedback

1.Claim calibration (R2, R3) We agree that several claims should be stated more precisely. In the revision, we will replace “physiological-pathological disentanglement” with “physiological-context-aware segmentation” and “zero-shot generalization” with “prior-guided zero-shot transfer.” We will also explicitly state the limited CD70 cohort size and the lack of center-disjoint evaluation as limitations, with broader validation left for future work. Thus, the submitted evidence supports improved context-aware multi-tracer segmentation, rather than unrestricted foundation-level generalization.

2.CT segmentation dependence and fairness (R1, R2) While relying on CT segmentation is a limitation we will note in the revision, ProMoE is architecturally designed for robustness against segmentation: (1) PE2 utilizes coarse patch-level organ occupancy rather than precise boundaries; (2) multiple experts whose joint similarities exceed a threshold are activated during training. This redundancy ensures that each expert is exposed to sufficiently diverse scenarios, thereby improving robustness to imperfect CT segmentation. Regarding comparison fairness, FiLM w/o MoE in Table 2 incorporates frozen CT backbone but excludes the MoE architecture. Its performance is inferior to external baselines such as nnUNet and SwinU, indicating that ProMoE’s gains do not stem from anatomical conditioning. Furthermore, under the K=12 setting, MLP-MoE performs worse than ProMoE, suggesting the improvements are not simply due to larger model capacity.

3.CD70 evaluation (R2, R3) For CD70, no images, labels, or results were used in the training stage. Only a fixed CD70 physiological prior, curated before evaluation (see response 4), was used at inference. We therefore redefine this as prior-guided zero-shot transfer. Its clinical significance lies in the fact that obtaining physiological distributions of new tracers is far more feasible than acquiring dense annotations. As R3 noted, this transferability assumes partial overlap between CD70 and seen tracers in normal uptake patterns. Fully new local uptake patterns may require extending the prototype space, which we will add as a limitation and future direction.

4.Reproducibility and atlas construction (R2, R3) Upon acceptance, we will release the code, configs, uptake atlas, and documentation needed to reproduce the reported models. The indicator 1(.) denotes non-zero support. S_Phy is non-negative, we empirically square it to emphasize stronger physiological consistency. We set θ = 1/K so that the sparsity threshold adapts to expert granularity. K=12 was selected by cross-validation on seen tracers. The atlas was LLM-drafted and then reviewed by two senior nuclear medicine experts. It is used as a fixed deterministic lookup table, so LLM randomness does not affect training or reproducibility.

5.Clinical aspects and why MoE (R3) We attempted to fuse encoded PET physiological text with image features using non-MoE architectures, but observed no performance gain. This is mainly because existing medical text encoders produce highly similar embeddings for structurally similar PET physiological pattern descriptions and cannot effectively resolve cross-tracer gradient conflicts discussed in Introduction. In contrast, ProMoE routes physiologically conflicting patches by expert prototypes to distinct experts, structurally alleviating this issue. Moreover, FDG dominates the absolute sample size in our dataset, its broader feature distribution fundamentally anchors the shared parameter space. Even with balanced batch sampling, single-branch models inevitably bias toward FDG-dominant patterns and fail on less frequent, more complex tracers. The results validate this: models w/o MoE fail on highly confounded tracers, whereas ProMoE achieves its most substantial gains on PSMA/FAPI. Conversely, baselines are already well-optimized for the dominant FDG distribution, explaining the marginal gains there.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

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

    The manuscript proposes a nice approach to segment lesions in PET/CT guided by anatomical and physiological knowledge inspired in the expert approach. The technical implementation is also parameter-effective while providing marginal advantages to the compared baselines. Yet, important aspects need to be explained in a rebuttal and modified in case the paper gets accepted. In particular, the claims made should be adapted to the obtained results (reviewer 2, criticism 4, reviewer 3, first paragraph of weaknesses), the dependance on CT segmentation should be discussed (reviewer 1, criticism 1 and reviewer 2, criticism 2), details should be given for the CD7 evaluation (reviewer 2, criticism 1, reviewer 3, optional comments), relevant aspects for reproducibility should be addressed (reviewer 2, criticism 3, reviewer 3, second paragraph of weaknesses), and clinical aspects of uptake patterns (reviewer 3) should be commented. Additionally, it is requested for the authors to make a statement if the code will be made available, as without it reproducibility is not really granted.

  • 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.

    In view of the post-rebuttal evaluation of the reviewers, I follow their recommendation and recommend approval. Still, I remain critical on the claims as I was on the first iteration and strongly encourage the authors to consider the post-rebuttal comments of reviewer 3.While I do not have means to enforce this, I hope the authors do not ignore the comment and include this in the limitation section.



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 rebuttal convinced the reviewers and they all agree for acceptance. A successful rebuttal period with addressing the majority of concerns.



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 proposed ProMoE, a prototype-driven Mixture-of-Experts framework with anatomical and physiological priors to integrate clinical reasoning into multi-tracer PET lesion segmentation. The major concerns from the previous round were properly addressed, including the CT segmentation dependence (patch level robustness is shown in Fig 3), details on the generalization test on an unseen dataset (CD70), and fairness and clinical aspects in MoE selection. My recommendation would also be acceptance.



back to top