List of Papers Browse by Subject Areas Author List
Abstract
Does every cancer patient truly need a complete diagnostic workup for accurate survival prediction? In multimodal clinical oncology, diagnostic modalities follow a clinically mandated order of escalating burden–from demographics collected at intake to genomic profiling requiring specialized tissue analysis. Current multimodal survival methods either assume all modalities are available or passively handle missing data, but none actively reason about whether acquiring the next modality is justified for a given patient along this ordered workflow. We formulate this as a sequential decision problem and propose SAGEAgent (Sequential Acquisition Guided by Experience), a self-evolving LLM-based clinical agent that decides which diagnostic modalities to acquire for each patient, balancing predictive accuracy against clinical invasiveness. SAGEAgent reasons about each patient’s evolving diagnostic state through clinical tools that translate numerical predictions into text, an episodic memory that retrieves similar past cases, and a semantic memory that accumulates reusable decision patterns from experience. Experiments on a glioma cohort combining TCGA-LGG, TCGA-GBM, and BraTS with four diagnostic modalities demonstrate that SAGEAgent achieves competitive survival prediction accuracy while reducing average acquisition burden by 55%.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0802_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Chongyu1117/SAGEAgent
Link to the Dataset(s)
N/A
BibTex
@InProceedings{QuCho_SAGEAgent_MICCAI2026,
author = { Qu, Chongyu AND Cui, Can AND Lu, Zhengyi AND Zhu, Junchao AND Yao, Tianyuan AND Guo, Junlin AND Xiong, Juming AND Zhu, Yanfan AND Yang, Yuechen AND Landman, Bennett A. AND Huo, Yuankai},
title = { { SAGEAgent: A Self-evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Optimizing the cancer diagnosis pathway.
- 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.
Novel methodology
- 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.
lack of external validation Lack of statistical analysis of model performance metrics.
- 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.
(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?
I am not convinced that the reported 55% reduction in patient burden sufficiently offsets the relatively small accuracy displayed when compared to existing solutions.
- 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.
Again, the proposed method is very interesting, and I see clear value in the reported 55% burden reduction. However, the absence of external validation represents a major limitation of the current study. That said, this issue could be reasonably addressed in a future publication through validation in independent cohorts. Overall, the potential clinical impact of the method is noteworthy and of substantial interest.
Review #2
- Please describe the contribution of the paper
This paper addresses key limitations of clinical multimodal glioma survival prediction. Existing static multimodal fusion methods generally require all diagnostic modalities to be collected before making a prediction, while ignoring substantial differences across examinations in economic cost, invasiveness, and diagnostic turnaround time, thereby imposing unnecessary clinical burden on many low-risk patients. Prior studies on dynamic modality acquisition are mostly based on reinforcement learning. Although they can reduce the number of acquired modalities, their decision process is a black box and lacks clinical interpretability and auditability, making them difficult to deploy in practice. To address these issues, the authors propose SAGEAgent, a fully decoupled two-module framework. The first module is a Transformer-based multimodal survival predictor with a calibrated uncertainty head, trained with modality dropout and partial-modality patient data so that it can handle common missing-modality scenarios and output both survival risk and calibrated uncertainty for any clinically valid modality prefix. The second module is an LLM-based decision agent with dual memory modules, which retrieves similar historical cases through episodic memory and stores updatable clinical decision rules through semantic memory, while strictly following the standard clinical diagnostic order to generate traceable natural-language reasoning. Based only on predictor outputs, similar historical cases, and fixed clinical rules, the agent dynamically decides whether to stop modality acquisition, thereby avoiding hallucination. On a mixed cohort of 962 glioma patients, SAGEAgent reduces average clinical burden by 42.7% compared with the full-modality static fusion baseline, with only a 0.013 drop in C-index that is not statistically significant, and improves C-index by 0.082 over a reinforcement-learning-based dynamic acquisition baseline while providing fully auditable and interpretable decision paths. The core innovation is that the work jointly incorporates clinical modality order constraints, multidimensional burden, and interpretability requirements into the optimization framework for multimodal prediction.
- 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 extends the optimization objectives of prior work beyond predictive accuracy by incorporating clinically important requirements, including modality acquisition order constraints, multidimensional clinical burden such as economic cost, invasiveness, and turnaround time, and decision auditability. This provides a new modeling direction for clinical multimodal decision support research. 2) The proposed fully decoupled predictor-agent architecture balances predictive stability and decision interpretability. The frozen Transformer-based multimodal predictor preserves the predictive stability of static fusion models, while the LLM-based decision agent with dual memory modules addresses the black-box nature of prior reinforcement-learning-based dynamic acquisition approaches. Its reasoning is based only on quantified predictions, similar historical cases, and standardized clinical rules, which makes the decision path traceable and reduces the risk of hallucination at the mechanism level. 3) The training and evaluation pipeline is more aligned with realistic clinical scenarios. The modality-dropout strategy combined with partial-modality patient data augmentation improves robustness to common missing-modality settings, and the MCDA framework enables a more comprehensive evaluation of multidimensional clinical burden than prior approaches that only use the number of acquired modalities as the evaluation metric. 4) The validation results on a public mixed glioma cohort show initial clinical application potential. The method reduces average clinical burden by 42.7% with only a non-significant 0.013 decrease in C-index, and the modular decoupled design allows rapid replacement of the multimodal predictor and decision rules without global retraining. The fully auditable decision pipeline may also reduce barriers to clinical compliance.
- 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 experimental comparison and generalization validation are insufficient. The method is validated only on a single-disease public glioma cohort and does not cover other cancer types or multi-center clinical settings. In addition, the set of comparison methods in dynamic modality acquisition is incomplete, since the experiments only compare against a basic reinforcement learning baseline and a full-modality static fusion baseline, without including more recent agent-based dynamic modality acquisition methods, making it difficult to fully establish the effectiveness of the SAGEAgent architecture or its cross-scenario generality. 2) The study lacks support from real clinical data. All experiments are based on retrospective offline analyses using public datasets, without prospective validation in real clinical workflows or verification on independent multi-center clinical cohorts. As a result, the actual clinical deployment effect and clinician acceptance remain to be validated. 3) There are unresolved gaps regarding training reliability and pathological validity. The lightweight Transformer-based multimodal survival predictor is trained with Cox survival loss and survival metrics under high-dimensional multimodal inputs and limited sample size, which introduces a substantial risk of overfitting, and the paper does not demonstrate at the principle level that the model truly learns core pathological features that are strongly related to survival outcomes. At the same time, similar-case retrieval in the episodic memory module is based only on vector similarity of input modality features and does not consider intrinsic heterogeneity in the pathological microenvironment, especially the tumor microenvironment. Therefore, feature-level similarity may differ substantially from true pathological similarity, which may reduce the clinical value of the retrieved cases. 4) The optimization design may not align with the underlying logic of clinical decision-making. Clinical diagnosis is a complex process that depends heavily on expert experience, and diagnostic accuracy, examination cost, and invasiveness are naturally in tension with one another. Integrating them into a single weighted objective may lead to performance degradation on both types of goals. The notes argue that a more appropriate design would treat diagnostic accuracy, examination cost, and invasiveness as three independent objectives and optimize them through a cooperative multi-agent game framework rather than the current weighted single-objective formulation. 5) Reproducibility and ablation validation are incomplete. The authors do not disclose the training data, parameters, or performance of the pretrained modality feature extractors, so third parties cannot independently verify the quality of multimodal feature encoding or fully reproduce the reported results. In addition, the ablation study does not include validation of modality contribution, and therefore cannot establish the actual contribution of different modalities such as genomics and imaging to the final predictions.
- 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?
The paper addresses an important problem in clinical multimodal survival prediction by moving beyond conventional static fusion and proposing a decoupled framework that separates survival prediction from modality acquisition decisions. The motivation is clear: existing static approaches are not robust to missing modalities and can impose unnecessary burden on low-risk patients, while reinforcement-learning-based dynamic acquisition methods are difficult to interpret and do not satisfy clinical compliance requirements. In this respect, the paper has several notable strengths, including a broader problem formulation that incorporates clinical order constraints, multidimensional burden, and auditability, a decoupled architecture that combines predictive stability with traceable decision-making, and a training and evaluation pipeline that is more relevant to realistic clinical settings through modality-dropout training and MCDA-based burden quantification. The reported results on a public glioma cohort are also encouraging, showing a substantial reduction in clinical burden with only a small and statistically non-significant decrease in predictive performance. However, the paper still has important limitations. The validation is restricted to a single-disease public cohort, the baseline comparisons are incomplete, and all evidence remains retrospective and offline without real clinical or multi-center verification. In addition, the reliability and pathological validity of the predictor and similarity retrieval strategy are not fully established, the weighted single-objective optimization design may not fully reflect the logic of real clinical decision-making, and the lack of disclosure for pretrained feature extractors together with incomplete ablations weakens reproducibility and interpretability. Overall, the work is meaningful and promising, but the current evidence is not sufficient to support acceptance.
- 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 design of an agent based on clinical examination decision-making is itself a novel and highly significant direction, particularly the rule-chain type of work implemented with agents. My concerns primarily lie in the completeness of the experiments, and the authors have also acknowledged the scarcity of data.
Review #3
- Please describe the contribution of the paper
The paper proposes SAGEAgent, a training-free LLM-based agent that sequentially decides whether to acquire the next diagnostic modality for a glioma patient along a clinically mandated ordering (demographics → radiology → pathology → genomics). The agent uses a frozen predictor plus a calibrated uncertainty head, clinical tools that verbalize numerical outputs, an episodic memory over past cases and the agent’s own decisions, and a semantic memory of natural-language rules evolved via periodic self-reflection. On a 170-patient complete-modality glioma subset (TCGA-LGG/GBM + BraTS) evaluated with nested 5×5 CV, SAGEAgent reportedly reduces acquisition burden by 55% while staying within 0.012 C-index of full-modality fusion.
- 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.Problem framing is genuinely novel and clinically well-motivated.
2.Architecture is coherent and well-grounded.
3.Interpretability is real, not performative.
4.Rule-discovery analysis (Table 3, Sec. 3.5) is the interesting empirical result. The finding that stopping rules converge in 1–2 reflection cycles and persist, while acquisition rules churn and often get deprecated, is a genuinely useful observation about what self-reflective agents can and cannot learn reliably.
5.Experimental protocol is rigorous for the scale.
- 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.SAGEAgent’s C-index of 0.813 is below Self-Attention Fusion (0.825), MMD (0.822), MCAT (0.817), Reflexion (0.823), and even the ablation’s Base LLM (0.827), showing that every component added reduces C-index relative to Base LLM, and the contribution is squarely on the accuracy–cost trade-off (captured by HV), not accuracy preservation.
2.No statistical significance testing. Standard deviations are ~0.04–0.05, and differences between the six top methods (0.812–0.825) are well within one std. Without paired tests across folds, claims like “competitive with full-modality” and “only method that achieves meaningful burden reduction without compromising accuracy” are not statistically supported.
3.Missing critical baseline: uncertainty-threshold stopping. The calibrated uncertainty head already achieves AUROC 0.908.A naive rule, “stop when u_t < τ”, would test whether the LLM, tools, and dual memory actually add value beyond the uncertainty estimator. Without this baseline, I cannot tell if the semantic rules are doing real work or just reformulating thresholds. Given that the discovered stopping rules all boil down to “uncertainty is low, stop,” this is a serious omission.
4.Comparison with LLM-based acquisition is weak. Only Reflexion is compared. Hager et al. (Nature Medicine 2024, ref [13]) and Nori et al. (ref [18]) are cited but not benchmarked, even though they are the most direct LLM-based-diagnosis competitors.
- 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 problem formulation is novel, the architecture is principled, and the rule-discovery analysis is a real contribution worth publishing. However, the central empirical claims are not cleanly supported.
- 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.
Statistical significance (W2): The results reframe the central claim defensibly: SAGEAgent shows no statistically significant C-index loss against any C>0.8 method (all 95% CIs contain 0, p≥0.30), while burden reduction and HV gain are highly significant (p<0.001). This converts the overclaim of “accuracy preservation” into a supportable trade-off claim. The framing in the submitted paper still overstates accuracy parity, but the underlying empirical contribution holds up.
Uncertainty-threshold baseline (W3): This was my most important concern, and the rebuttal answers it directly. At τ=0.3/0.5/0.7, the threshold-only policy yields C-index 0.77/0.74/0.73 versus SAGEAgent’s 0.813 — a 0.04–0.08 gap that demonstrates the LLM, tools, and dual memory do contribute beyond the calibrated uncertainty estimator.The comparison is convincing on its face and consistent with the paper’s narrative about rule discovery.
LLM-based acquisition baselines (W4): The justification that Hager et al. and Nori et al. operate end-to-end on raw clinical text rather than pre-extracted multimodal features is acceptable, though an adapted comparison would have strengthened the paper.
In conclusion, the rebuttal resolves the substantive empirical concerns and the rule-discovery contribution remains the strength of the paper. Therefore, my opinion is stated as Accept.
Author Feedback
We thank the reviewers for their feedback and for recognizing that “the work is meaningful and promising”-R4 and “rule-discovery analysis is a real contribution worth publishing”-R5. Common concerns:
1) Statistical significance testing (R3:W2, R5:W2). Paired bootstrap (1000 reps, 170 patients) on ΔC-index, ΔBurden, and ΔHV (SAGEAgent vs comparator) for every Table 1/2 method with C > 0.8:
ΔC-index: every 95% CI contains 0 (all p ≥ 0.30). Concat Δ=−0.001 [−0.040,+0.035]. MCAT −0.005 [−0.037,+0.030]. MMD −0.010 [−0.053,+0.029]. Self-Att −0.013 [−0.050,+0.027]. Reflexion −0.011 [−0.047,+0.025]. Base LLM −0.014 [−0.050,+0.022].
ΔBurden: every 95% CI strictly below 0 (all p < 0.001). vs B=1 methods Δ=−0.549 [−0.589,−0.505]. Reflexion −0.420 [−0.473,−0.364]. Base LLM −0.346 [−0.395,−0.295].
ΔHV: every 95% CI strictly above 0 (all p < 0.001). vs static fusion Δ=+0.173 [+0.146,+0.198]. Reflexion +0.131 [+0.104,+0.158]. Base LLM +0.106 [+0.083,+0.129].
SAGEAgent has no significant C-index loss but significantly lower burden and higher HV than every C > 0.8 method. RL excluded due to substantial accuracy drop.
2) External validation (R3:W1, R4:W1). To our knowledge, TCGA-LGG/GBM + BraTS is the only public cancer cohort where 170 patients carry all four modalities. These 170 form our evaluation set, while the 792 partial-modality patients serve as predictor augmentation via modality dropout. Representative alternatives fall short on modality coverage: CPTAC-GBM (200 pathology, 70 radiology), PORPOISE (TCGA pan-cancer, n=5720, pathology and genomics only, no radiology). Cross-cancer and multi-center evaluation are listed as future work (Sec. 4) once matching public multimodal cohorts become available.
3) LLM agentic baselines (R4:W1, R5:W4). Our task uses pre-extracted multimodal features and requires a multimodal predictor. Reflexion and SAGEAgent operate above this predictor via text-translated signals, decoupling agent reasoning from input format. Hager et al. (Nat Med 2024) and Nori et al. (2025) operate end-to-end on raw clinical text with no feature-level multimodal pipeline, and cannot ingest our setting without substantial redesign.
R3: 1) See Common #2.2) See Common #1.3) The C-index difference is not statistically significant (Common #1, all p ≥ 0.30, CIs contain 0) while the 55% burden reduction is significant (p < 0.001).
R4: 1) See Common #2 (validation) and #3 (agent baselines). 2) TCGA-LGG/GBM and BraTS are public real clinical data (biopsy specimens, follow-up, multi-center MRI), supporting fair comparison and reproducibility. Prospective validation requires multi-year IRB-approved studies beyond our scope [Roberts et al., Nat Mach Intell 2021]. 3) Overfitting is mitigated by a 458K-param predictor, 50% modality dropout, 792-patient augmentation (Sec 3.1, 3.2), and nested 5×5 CV. Retrieval uses pre-extracted features [Chen et al., Cancer Cell 2022] rather than raw images, since per-query encoding of WSI and MRI would be computationally costly. 4) Our LLM is frozen (Sec 2.3) and makes binary acquire/stop decisions, with reward used only for memory ranking and rule discovery, not gradient-optimization of a weighted scalar. A cooperative multi-agent game with independent objectives is a promising future direction. 5) Feature extractor uses default settings from [Cui et al., MICCAI 2022], pre-extracted features will be released for reproducibility. Prompts, decoding params, seeds, and FAISS details are in the released code. Fixed-stopping marginal gains are rad +0.10, path +0.02, gen +0.001, supporting SAGEAgent’s skip pattern for pathology and genomics.
R5: 1)-2) See Common #1.3) Uncertainty-threshold stopping at τ=0.3/0.5/0.7 yields C-index 0.77/0.74/0.73 and burden 0.23/0.13/0.08, sacrificing 0.04 to 0.08 C-index compared to SAGEAgent (0.813), indicating that the LLM, tools, and dual memory add value beyond the uncertainty estimator. 4) See Common #3. Other correction: Qwen citation will be corrected.
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.
(1) The authors should provide paired statistical tests or bootstrap confidence intervals for C-index, burden, and hypervolume when comparing SAGEAgent with Self-Attention Fusion, MMD, MCAT, Reflexion, and Base LLM. In particular, the claim that SAGEAgent is “within 0.012 C-index” of full-modality fusion should be statistically supported, preferably with a predefined non-inferiority margin; otherwise, the claim of “no accuracy compromise” should be softened. (2) The authors should add an uncertainty-threshold stopping baseline. (3) The authors should clarify why Base LLM achieves a higher C-index than SAGEAgent, and explicitly present SAGEAgent as improving the accuracy–burden trade-off rather than pure predictive accuracy. (4)The authors should either include adapted LLM-based acquisition baselines or justify why prior LLM diagnosis methods such as Hager et al. and Nori et al. cannot be directly benchmarked. (5)The authors should clarify that the reported evaluation is based on 170 complete-modality patients, while the remaining 792 partial-modality patients are used for training augmentation. (6)More implementation and reproducibility details should be provided, including feature extractors, prompts, decoding parameters, random seeds, threshold tuning, FAISS retrieval features, reflection prompts, and the Qwen2.5/Qwen3 citation issue.
- 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.
All reviewers who provided post-rebuttal opinions recommend acceptance. The rebuttal adequately addresses the main empirical concerns, including statistical support for the accuracy–burden trade-off, the uncertainty-threshold stopping baseline, and the role of the LLM/tools/dual-memory components. Reviewers also recognize the clinical relevance and novelty of agent-based clinical examination decision-making. Remaining concerns, especially the lack of external validation, limited generalizability, and data scarcity, should be clearly discussed as limitations, but they do not outweigh the paper’s methodological and clinical interest. The authors should provide the full open-source implementation to improve the reproducibility and broader impact of the paper.
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.
Congratulations!
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 paper truly introduces a valuable contribution to the state-of-the-art. While deeper evaluations are needed towards clinical translation, I agree with the reviewers that the current version with the modifications resulting from the rebuttal phase make it worth presenting at MICCAI.