List of Papers Browse by Subject Areas Author List
Abstract
Multi-tumor segmentation is important for early cancer detection and allows radiologists to visualize, verify, and understand AI predictions. However, tumor segmentation masks are expensive, time-consuming, and unavailable for many tumor types in public data. Instead, hospitals have vast, readily-available data that can guide segmentation: radiology reports and longitudinal images with multiple contrast phases. We use this readily-available data to substitute for tumor masks in training AI for tumor segmentation. To this end, we propose a new architecture, RT-Super. It has a teacher network, which analyzes the patient’s longitudinal images and reports to create high-quality tumor masks. These masks train a student network, which sees a single image and no report. In inference, when longitudinal images and reports are unavailable, we use the student. RT-Super uses a new CNN-Transformer architecture and novel Consistency Losses that exploit tumors location consistency across longitudinal images. We train RT-Super to segment esophagus, uterus and spleen tumors, which have few or no public masks. Even without training masks, RT-Super can segment these tumors and surpass public AI models. Overall, we demonstrate that learning from longitudinal images, multi-phase images, and reports can overcome mask scarcity and advance multi-cancer detection and segmentation. Code: https://github.com/MrGiovanni/RT-Super
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4075_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MrGiovanni/RT-Super
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BasPed_RTSuper_MICCAI2026,
author = { Bassi, Pedro R. A. S. AND Li, Wenxuan AND Gu, Hanxue AND Chen, Jieneng AND Zhou, Xinze AND Zhu, Zheren AND Er, Sezgin AND Hamamci, Ibrahim E. AND Menze, Bjoern H. AND Akan, Gulhan E. AND Wang, Kang AND Yang, Yang AND Yuille, Alan L. AND Zhou, Zongwei},
title = { { RT-Super: Learning Tumor Segmentation from Longitudinal Images and Reports } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes TR-Distillation, a teacher-student/self-distillation framework for multi-cancer tumor detection and segmentation under severe mask scarcity. The teacher uses radiology/pathology report attributes together with longitudinal and multi-phase CT scans, while the student is trained to infer from a single image without reports, which is the clinically practical deployment setting. The work also introduces a report-aware transformer that generates convolutional kernels, inter-image cross-attention across scans, and a longitudinal consistency loss. In addition, the paper assembles a sizable internal dataset of 5,000 CT scans with reports focused on tumor types that lack public segmentation masks: esophagus, spleen, and uterus tumors.
- 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 highly relevant and practical problem: most hospitals have abundant reports and serial imaging but very limited voxel-wise tumor masks. The proposed training/inference asymmetry is therefore well motivated clinically: rich context is used only at training time, while inference remains feasible from a single scan. 2.The paper presents strong internal empirical results. In Table 1, the full TR-Distillation model achieves the best overall detection F1 among methods trained on the authors’ dataset and clearly outperforms the public baselines reported in the paper. The ablation labeled “non-longitudinal” also suggests that the longitudinal/multi-phase training signal contributes measurably beyond report supervision alone. 3.The work has clear translational value because it targets tumor types for which public segmentation resources are scarce, and it explicitly studies both the no-mask and few-mask regimes. That makes the contribution more meaningful than a purely incremental accuracy improvement on an already saturated benchmark.
- 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 main novelty is somewhat integrative rather than foundational. Report-based supervision for segmentation is already present in R-Super [2], report/image representation learning is already explored through CLIP-style and vision-language approaches [7,18], the student backbone is MedFormer [11], dynamic kernel selection is explicitly inspired by Dynamic Convolution [10], and registration is built on UniGradIcon [20]. The paper’s real novelty is the combination of these ideas with longitudinal/multi-phase information, which is meaningful, but the manuscript occasionally presents the advance more broadly than the evidence supports. 2.The ablation study is not deep enough to isolate where the gains actually come from. The paper shows a non-longitudinal ablation, but it does not disentangle the impact of: report-aware kernel generation, inter-image cross-attention, consistency loss, LLM-extracted attributes, and the use of pathology information. Without such decomposition, it is difficult to judge which proposed component is essential and which is incidental. 3.The evaluation section lacks critical detail, especially for segmentation. The manuscript reports Dice scores, but the description of the test set emphasizes detection labels from reports/pathology and does not clearly explain how the segmentation ground truth on the test cohort was obtained, how many test masks exist, who annotated them, or whether there was any quality control or inter-rater assessment. This is an important omission for a segmentation paper. 4.The paper is also limited by single-system internal validation. All training and testing data come from the same hospital system/state, so the claims of robustness and broad clinical applicability remain unproven. Given that reporting style, scanner protocols, contrast timing, and follow-up practices can vary substantially across institutions, an external validation set would materially strengthen the work. 5.Some of the public-model comparisons are not fully apples-to-apples. For example, Merlin and MedGemma are used through report generation for tumor detection rather than through a segmentation setting aligned with the proposed task. These comparisons are still interesting, but they are less convincing than the same-data comparisons against R-Super, segmentation, CLIP, and related baselines.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
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?
My overall score is driven by the combination of strong practical motivation, a well-designed and clinically plausible training/inference setup, and convincing internal gains over the baselines presented in the manuscript. I find the idea of using longitudinal scans and reports as training-only supervision for a single-image student to be both useful and timely. At the same time, I am holding the paper back from a stronger accept because the novelty is partly integrative, the ablations are not sufficiently granular, and the experimental section leaves important questions unanswered about segmentation ground truth and external generalization. On balance, I think the paper is above the bar, but only marginally so, and I would expect the rebuttal to clarify the evaluation protocol and novelty positioning.
- 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.
I am changing my recommendation to Reject after the rebuttal. The author response contains multiple new/additional experimental results, which are explicitly not allowed by the MICCAI rebuttal rules.
Review #2
- Please describe the contribution of the paper
This paper presents TR-Distillation, a self-distillation framework for multi-tumor CT segmentation that replaces expensive voxel-level tumor masks with readily available hospital data—radiology reports, pathology reports, longitudinal images, and multi-phase contrast images. The teacher network is a hybrid CNN-Transformer that ingests all available patient data (reports and longitudinal/multi-phase scans) to produce pseudo-masks via a report-aware cross-attention mechanism and a dynamic convolutional kernel bank (soft mixture-of-experts style). The student network, which is embedded as a subnetwork of the teacher (self-distillation), operates on a single image without reports at inference. A novel Consistency Loss enforces spatial alignment of tumor predictions across longitudinal scans. The framework is trained on 5,000 CT scans with reports covering esophagus, spleen, and uterus tumors—three types that lack public segmentation masks—and demonstrates substantial improvements over public AI models (MedGemma, Merlin, ULS) and several weakly/semi-supervised 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.Clinically grounded motivation and problem formulation. The observation that radiologists routinely consult longitudinal scans, multi-phase images, and prior reports when delineating tumors is compelling. The paper translates this clinical workflow into a principled teacher-student framework where the teacher mimics the radiologist’s information-rich setting and the student mimics the information-poor setting at first diagnosis. The focus on esophagus, spleen, and uterus tumors—which currently lack public segmentation masks—addresses a genuine gap in the field.
2.Novel and technically interesting architecture. The report-aware transformer that dynamically generates convolutional kernels is a creative design. Specifically, the use of a learnable bank of M = 16 candidate 3×3×3 kernels combined via softmax-weighted mixture (Sec. 2.1, page 5) avoids the computational cost of directly generating large kernels, while still allowing report-conditioned feature refinement. The self-distillation design—where the student is physically embedded within the teacher—is efficient and avoids maintaining two separate networks.
3.Comprehensive comparison against public models. The evaluation includes public foundation models (MedGemma, Merlin) and a dedicated lesion segmentation model (ULS), in addition to multiple training paradigms (CLIP, multi-task learning, Models Genesis, Report Supervision, standard supervised segmentation). The +49% average F1-score improvement over the best public model (Table 1) is substantial.
- 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.Vague and ambiguous dataset description. The dataset description in Section 3 is insufficiently precise for a methodological paper and raises several unresolved questions. The text states: “5,000 CT scans paired radiology reports and pathology reports, including esophagus, spleen and uterus tumors, 100 longitudinal CT scans.” It is unclear whether the 100 longitudinal CT scans are a subset of the 5,000 or an additional set, yielding 5,100 total scans. Similarly, the paper mentions “100 multi-phase scans” on page 2 but does not clarify their relationship to the 5,000.The two training variants are described as: TR-Distillation† using 5,000 image–report pairs, and TR-Distillation using 5,000 image–report pairs plus 100 image-mask pairs. But how exactly are the 100 longitudinal scans incorporated during training? Are they always included in both variants? How many of the 5,000 scans contain tumors versus being non-tumor controls? What is the per-organ breakdown of the training set? Without these details, it is difficult to assess the experimental setup or reproduce the work.
2.Unclear and potentially unfair baseline training settings. Closely related to the vague dataset description, the training conditions for baseline methods are not clearly specified, making it difficult to assess whether comparisons are fair. Table 1 marks each method’s use of “longi.” (longitudinal), “report,” and “mask” columns, but the actual number of training samples used by each baseline is never stated explicitly. For instance, it appears from context that methods marked with “report” and “mask” (e.g., R-Super, CLIP, Multi-task learning) may use reports and 100 masks, but do they all have access to the same 5,000 image–report pairs? Methods like Models Genesis and the supervised segmentation baseline appear to use only 100 masks without reports—but do they also train on the remaining ~4,900 unlabeled images in any capacity (e.g., self-supervised pre-training)? The classification baseline is marked only with “report” but its training procedure is barely described. Without explicit, per-baseline specification of how many images, reports, and masks each method uses, the reader cannot determine whether performance differences are attributable to the proposed architecture or to differences in data volume and supervision type. This is especially concerning because TR-Distillation leverages 5,000 reports plus longitudinal data plus multi-phase data—potentially a much richer supervisory signal than what baselines receive. The authors should provide a clear table specifying the exact training data (number of images, reports, masks, longitudinal pairs) used by each method.
3.Dice scores reported only in the table caption, not in the body text. The table caption for Table 1 reports Dice scores: “TR-Distillation: 35 average; 37 spleen; 68 esophagus; 71 uterus. TR-Distillation†: 17 average; 7 spleen; 27 esophagus; 25 uterus. R-Super (best baseline): 30 average, 36 spleen, 21 esophagus, 34 uterus.” However, these Dice scores do not appear anywhere in the body text of Section 3 or elsewhere in the paper. They are also absent from Table 1 itself, which only shows sensitivity (Se), specificity (Sp), and F1-score columns. This is problematic for several reasons. Dice is the standard metric for segmentation quality, and yet it is relegated to a caption with no discussion, no statistical analysis, and no comparison against all baselines. The Dice values themselves are notably low (e.g., 35 average for TR-Distillation, 7 spleen Dice for TR-Distillation†), which warrants discussion. Are these values computed only on true-positive cases? Over all cases including non-tumor controls? The omission of Dice from both the main table and the body text is a significant gap for a segmentation paper.
4.Insufficient ablation study. The paper presents a single ablation that removes longitudinal/multi-phase training entirely (Table 1, bottom row). For an architecture with at least five novel or adapted components—(i) the Consistency Loss, (ii) report cross-attention, (iii) inter-image cross-attention, (iv) dynamic kernel generation via the soft MoE bank, and (v) soft vs. hard distillation targets—a single ablation is inadequate. Without isolating each component, we cannot determine whether the Consistency Loss actually helps, whether the dynamic kernel bank outperforms standard convolutions, or whether soft distillation targets contribute meaningfully. A systematic ablation table (removing one component at a time) is standard for multi-component methodological papers and is essential here.
W5.Teacher performance is never reported. The entire framework rests on the premise that the teacher, with access to reports and longitudinal data, produces high-quality pseudo-masks that effectively train the student. Yet the teacher’s segmentation performance is never evaluated anywhere in the paper. This is a significant omission for two reasons: (1) it is impossible to assess the quality of the pseudo-masks that drive learning, and (2) the “distillation gap” (teacher performance minus student performance) remains unknown. If the teacher’s masks are poor, the student cannot be expected to learn well; if the teacher significantly outperforms the student, there may be room for improvement in the distillation mechanism. Reporting teacher Dice scores alongside student scores is essential for validating the core mechanism.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
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 a well-motivated and important problem—learning tumor segmentation from readily available hospital data rather than expensive masks—and proposes a technically interesting architecture. However, several fundamental issues prevent confident acceptance: (1) the dataset description is too vague to fully understand the experimental setup; (2) the training conditions for baselines are unclear, raising concerns about fairness of comparisons; (3) Dice scores—the primary segmentation metric—appear only in a table caption with no discussion or full reporting; (4) the ablation study is too coarse to attribute performance to specific components of a multi-component system; and (5) the teacher’s performance is never reported, leaving the core distillation mechanism unvalidated. Addressing these issues—by providing precise dataset and training details, equalizing and documenting data access across baselines, fully reporting and discussing Dice scores, adding component-wise ablations, and reporting teacher performance—would substantially strengthen the paper.
- 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.
Although I still have concerns about the paper’s clarity, its motivation and methodology are quite attractive, and my major concerns about the dataset description and ablation studies have been addressed. So I move towards an accept.
Review #3
- Please describe the contribution of the paper
In clinical practice, diagnostic and treatment decisions are typically based on longitudinal imaging, culminating in radiology and pathology reports rather than the availability of precise 3D segmentation masks. This paper demonstrates that such realistically available outputs can be effectively leveraged to train improved models, compared to approaches that do not utilize this information.
- 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 is clearly motivated by real clinical practice. It designs experiments that demonstrate the value of leveraging additional data sources that are often more readily available than clean segmentation masks or classification labels. This is valuable for guiding data curation strategies, as the proposed setting better reflects real clinical conditions, including longitudinal imaging, radiology reports, and pathology reports at the patient level. 2.The paper provides clear explanations at each step and is transparent about its core technical components, such as the loss functions used, how ground-truth labels are handled when available, and the use of MedFormer as the backbone. The technical details are also well motivated, including what information is extracted from existing reports and how it is incorporated into the model.
- 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 overall Dice scores remain relatively low (ranging from 30–70 in Table 1), which raises concerns about practical performance. In addition, the presentation is somewhat unclear, as the Dice metric is only described in the table caption rather than being explicitly labeled within the table itself. 2.The segmentation baselines and comparisons appear limited. The evaluation includes only three methods, two of which are based on TR-Distillation, resulting in a lack of diversity in comparison. It would strengthen the paper to include more established baselines, such as MedFormer and nnU-Net trained with the full set of 100 labeled samples.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
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 concerns raised in the weaknesses do not diminish the overall value of the paper. However, addressing and clarifying these points would strengthen the work and could justify a higher rating. The paper is particularly compelling in terms of its motivation and the results it demonstrates; the narrative is well connected and largely convincing. Moreover, it offers an insightful perspective on how to mitigate the challenges of curating segmentation masks in real-world clinical settings.
- 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.
I will keep my decision as accept, since the authors added additional results using nn-UNet, which was previously missing as an important segmentation baseline. They also improved the presentation of the results by moving the Dice scores into the table, making the comparisons clearer and more effective.
Author Feedback
Thanks for the 2 Weak Accepts, and for saying our study is “novel and technically interesting”, “well motivated”, “useful and timely”.
Ablations (TR-Distill trained w/ reports and masks) AUC: Esophagus, Spleen, Uterus, avg No inter-image cross-att.:-1.1,-0.9,-0.8,-0.9 No dynamic kernel gen.:-2.5,-2.3,-0.6,-1.8 No hard distill:-2.2,-1.4,-0.9,-1.5 No report cross-att.:-4.6,+0.2,-1.6,-2.0 No soft distill:-1.1,-1.9,-0.4,-1.1 No consistency loss:-2.8,-1.5,-1.3,-1.9 No LLM extracted tumor size:-2.1,-0.3,-1.7,-1.4
DSC details: now in the table (was in caption), for all baselines, with std. DSC computed on true positives. DSC is high for esophagus and uterus (69, 72; over public datasets as PanTS). It is low for spleen (40) due to more small spleen tumors in test set (39% vs 24 eso/9 ute). DSC: spleen esoph. uterus Ours: 40 +/- 20, 69 +/- 27, 72 +/- 28 nnU-Net (new): 35 +/- 23, 37 +/- 29, 45 +/- 32
R1 1.Novelty: We introduce a novel training strategy to learn tumor segmentation in the absence/scarcity of masks: a teacher uses longitudinal images and their reports to train a student that sees a single image and no report. No previous work used this framework. We agree it leverages existing methods (cited) as components (eg, UniGradIcon for registration), besides fully new components (eg, consistency loss). We have now clarified this better in the camera ready.
- Ablations: 7 new ablations included at the top, the pathology ablation will be in the camera ready due to time constraints.
3.Test masks: drawn by 31 radiologists (2 to 19 experience years); 38 spleen, 181 esophagus, 105 uterus. inter-reader DSC (2 readers, 2 / 8 years of experience): 66, 50, 62.The tumor types in our dataset are more difficult to detect than common types (e.g., kidney), making inter-reader and model performance lower.
4.External eval: is now included. It includes only detection metrics, but we will partially annotate the data for the camera ready to include DSC. Notably, the new data comes from Turkey, and the model was trained in the USA. Ours (AUC eso/spl/ute): 85/81/89 nnU-Net: 68/75/83; CTs: 189/347/127. 5.Baselines: We agree that comparison to other segmentation models trained on our dataset are the most meaningful, we included 6 of such baselines in the paper, and we also added the nnU-Net now (see top).
R2 1.Dataset details: The longitudinal & multi-phase scans are subsets of the full dataset (5K). So, longitudinal scans were included in the training datasets of both TR-Distillation variants; 77% of the training dataset contains tumors (9% esophagus, 46% spleen, 22% uterus).
2.Baseline training: All baselines we trained were trained on the same data as TR-Distillation, so all baselines that used reports had access to the same 5K image-report pairs as our model, including the longitudinal and multi-phase cases. ours, R-Super, CLIP, MTL (use reports&masks): 4.9K image–reports + 100 image-masks. Models Genesis: Self-supervised pre-training on 5K images, fine-tuning for segmentation on 100 image-masks. Segmentation (and nnU-Net, new): 100 image-mask pairs only. Self-supervised pretraining covered above. Classification: multi-class tumor labels (e.g., uterus tumor present vs absent) extracted from the reports by LLM and used to train a classifier on all dataset (5K)
3.Please see the DSC discussion on top.
4.Ablation: we added 7 ablations, including all you requested (top)
5.Teacher performance: teacher’s DSC is +2.9% better than the student.
R3
1.DSC moved to table, and discussion on top of this rebuttal.
2.Segmentation baselines: we compare to 10 baselines, 6 do segmentation and are trained on our data: CLIP (CT-Report pretrain, CT-Mask segm. fine-tune), R-Super (segmenter), Multi-task learning (classification+segm.), Models Genesis (SSL pretrain, segm. fine-tune), and “segmentation” (MedFormer trained on all CT-masks, as you suggested). We now added nnU-Net, see top.
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.
Please answer the concerns raised by the three reviewers in the rebuttal, especially on training settings, ablation study, segmentation performance and more baselines, and dataset description.
- 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 presents a valuable method to learn tumor segmentation from longitudinal CT and radiology reports. The authors’ rebuttal further added important baselines, ablation studies, dataset, method and experimental clarifications. Two reviewers recommended acceptance. Although R1 pointed out that MICCAI does not allow new results in rebuttal, I think the paper represents a contribution that is attractive to the MICCAI community.
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 core idea of teacher-student distillation to address the problem of severe label scarcity for segmentation was seen as interesting by all three reviewers. The authors addressed most of reviewers’ concerns although concerns regarding the clarity of the dataset and completeness of experiments. Authors added new experiments, which are not allowed according to the MICCAI rules. It would be better if the authors kept the additional experiments to bare minimum needed (e.g. nnU-Net) and instead address concerns in the discussion.
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.
After reviewing the comments and the authors’ responses, I find that most concerns have been addressed. Overall, the paper is in good shape for acceptance. I recommend accepting the submission, while encouraging the authors to incorporate the reviewers’ suggestions carefully when preparing the final version.