List of Papers Browse by Subject Areas Author List
Abstract
While vision-language models dominate medical representation learning, unstructured text lacks the dense, quantitative diagnostic phenotypes inherent in structured clinical tables. However, existing multimodal pre-training methods underutilize this potential due to semantic-agnostic designs that treat tabular inputs as flat vectors and employ unstable continuous regression objectives. To overcome this, we propose a novel semantic-aware framework explicitly modeling the intrinsic two-dimensional structure of tabular data. First, addressing the inter-feature hierarchy of varying diagnostic importance, we introduce Importance-Aware Adaptive Masking to construct a label-free curriculum prioritizing salient features. Second, addressing the intra-feature continuity-discreteness duality, we propose a Soft-Label Discretized Module that replaces unstable numerical regression with stable distribution matching, thereby mathematically preserving ordinal relationships. Extensive experiments across large-scale dermatology (SLICE-3D, HOP) and ophthalmology (EyePACS) datasets establish a new state-of-the-art (SOTA), demonstrating exceptional robustness and cross-domain generalizability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0019_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Ethan-ysliu/AID
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuYin_Unlocking_MICCAI2026,
author = { Liu, Yingsheng AND Li, Haiming AND Zhu, Jingmin AND Sun, Jiajun AND Mar, Victoria AND Janda, Monika AND Soyer, H. Peter AND Ge, Zongyuan AND Yu, Zhen},
title = { { Unlocking the Power of Medical Tabular Data via Semantic-Aware Multimodal Pre-training } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes AID (Adaptive Importance-guided Discretized reconstruction), a semantic-aware multimodal pre-training framework for joint image-tabular representation learning in medical imaging. The main contributions are threefold.
First, the authors identify and formalize a two-dimensional hierarchical structure inherent to medical tabular data: an inter-feature dimension capturing varying diagnostic importance across clinical attributes, and an intra-feature dimension capturing the continuity-discreteness duality within individual features. This conceptual framework motivates the entire design.
Second, an Importance-Aware Adaptive Masking module is introduced to construct a label-free curriculum. Feature significance is estimated offline using a frozen TabPFN v2 meta-prior driven by PCA-derived pseudo-labels, avoiding label leakage while capturing dataset-specific feature interactions. Masking rates are then modulated proportionally to feature importance, compelling the model to rely on cross-modal visual evidence for reconstructing clinically salient features.
Third, a Soft-Label Discretized Module replaces unstable continuous regression with a stable distribution-matching objective. A triangle kernel allocates probability mass across adjacent quantile bins, mathematically preserving ordinal relationships of clinical measurements while avoiding the boundary-discontinuity artifacts of hard-label discretization.
The combined pre-training objective unifies ITC, ITM, and the novel DR loss. Evaluated on large-scale dermatology (SLICE-3D, HOP) and ophthalmology (EyePACS) datasets, AID establishes a new state-of-the-art, with particularly strong out-of-domain generalization (OOD AUC 0.944 vs. 0.911 for the strongest baseline), demonstrating robustness to geographic distribution shift and cross-specialty transferability.
- 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 problem formulation. The two-dimensional hierarchical structure hypothesis (inter-feature importance and intra-feature continuity-discreteness duality) is not an abstract theoretical construct but directly reflects how clinicians reason about structured patient data. This conceptual framing is the paper’s most original contribution and provides a coherent motivation for both proposed modules.
2.Strong out-of-domain generalization. The geographically isolated OOD test set on SLICE-3D is a rigorous evaluation design rarely seen in multimodal pre-training papers. Achieving an OOD AUC of 0.944 versus 0.911 for the strongest baseline (TIP) is a substantial and clinically meaningful margin. Furthermore, the linear probe AUC of 0.984 surpassing TIP’s full fine-tuning AUC of 0.971 demonstrates exceptional representation quality.
3.Thorough and honest ablation study. Table 2 systematically evaluates not only the presence or absence of each module, but also compares multiple alternative design choices for discretization (equal-width bins, hard-label quantile, Gaussian kernel, triangle kernel). The finding that the simpler triangle kernel outperforms the heavier Gaussian kernel is a practically useful and honestly reported result.
4.Label-free importance estimation is well-designed. Using PCA pseudo-labels to drive a frozen TabPFN v2 meta-prior for feature importance extraction is a principled approach that avoids label leakage while capturing dataset-specific feature interactions. This design choice reflects careful thinking about the practical constraints of medical pre-training.
5.Cross-specialty validation strengthens generalizability claims. Applying the framework to EyePACS (ophthalmology) in addition to dermatology datasets provides meaningful evidence that the two-dimensional structural hypothesis holds across medical domains, not just within the training specialty.
6.Qualitative analysis corroborates quantitative gains. The t-SNE visualizations and attention heatmaps in Figure 3 provide interpretable evidence that the soft-label objective produces semantically organized latent clusters and clinically meaningful attention distributions, supporting the mechanistic claims of the paper.
- 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.Sensitivity to the number of discretization bins (B) is not evaluated. The framework uses B=50 quantile bins, but this hyperparameter directly controls the granularity of the soft-label distributions and the smoothness of ordinal preservation. No ablation study is provided for B, and it is unclear whether performance is robust to this choice across datasets with different feature distributions. A brief sensitivity analysis in the supplementary would address this concern.
2.Computational cost of the importance estimation step is not reported. The paper states that TabPFN v2 feature importance extraction is performed strictly offline prior to pre-training, but does not report the wall-clock time or memory requirements for this step. For practitioners working with large-scale datasets (e. g. , SLICE-3D with 400,000 images), this information is relevant for assessing the practical overhead of adopting the framework.
3.The three-head ensemble mechanism is insufficiently described. Section 3.1 mentions that downstream evaluation uses a “three-head ensemble mechanism that averages distinct classification logits,” but does not specify what the three heads are or how they differ. No ablation is provided to justify this design choice over a single-head classifier.
4.The high maximum masking rate may raise stability concerns. With rbase=0.1 and alpha=rmax=0.7, the most important features can receive a masking rate of up to 0.8.While the paper reports stable training, no analysis or discussion is provided regarding potential optimization instability at this extreme masking rate, particularly for datasets with fewer training samples.
5.The HOP private dataset limits full reproducibility. HOP is a private institutional dataset and no detailed statistics (class distribution, image acquisition protocol, patient demographics) are provided. Since HOP is used as a primary external validation set, the absence of this information makes it difficult to assess the significance of the reported gains on this dataset.
6.The paper title is suppressed in the submitted version (“Title Suppressed Due to Excessive Length”), indicati
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The primary factors driving a strong accept recommendation are as follows.
The paper addresses a genuine and clinically relevant problem. Multimodal pre-training for image-tabular data is an important direction for medical AI, and the identified limitation of semantic-agnostic designs (uniform masking and continuous regression) is well-motivated and supported by both theoretical argument and empirical evidence.
The performance gains are substantial and rigorous. The OOD geographic holdout evaluation on SLICE-3D is a particularly strong experimental design, and the margin over the strongest baseline (TIP) is large enough to be clinically meaningful rather than a marginal improvement. The fact that linear probing outperforms full fine-tuning of the prior state-of-the-art is a compelling indicator of representation quality.
The ablation study is thorough and the conclusions are reliable. The systematic comparison of discretization strategies, including the honest reporting that the simpler triangle kernel outperforms the Gaussian kernel, demonstrates scientific rigor.
Cross-specialty validation on EyePACS provides meaningful evidence of generalizability beyond a single domain, which is important for a method claiming broad applicability to medical tabular data.
The identified weaknesses (missing bin sensitivity analysis, insufficient description of the ensemble mechanism, private HOP dataset) are real but do not undermine the core contributions. They are straightforwardly addressable in the camera-ready version and do not cast doubt on the validity of the reported results.
In summary, this paper makes a clear methodological contribution with strong empirical support, directly relevant to the MICCAI community. It meets the bar for acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper presents an enhanced multimodal pre-training framework that integrates medical tabular data with imaging by addressing key limitations in the TIP architecture (Du et al. 2024). The authors specifically target the weaknesses of semantic-agnostic random masking and the overfitting tendencies of linear reconstruction methods. To resolve these, they introduce two technical innovations: Label-Free Feature Importance Extraction, which weights diagnostic relevance without supervised labels, and a Soft-Label Discretization Module to stabilize continuous feature reconstruction. The method consistently outperforms state-of-the-art baselines across three clinical datasets, yielding a 2% to 4% improvement in AUROC. Comprehensive ablation studies and qualitative analyses further confirm that these targeted enhancements produce more robust representations for multimodal medical tasks.
- 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.
- Critical analysis: The authors provide a precise critique of the TIP architecture’s limitations, followed by a well-justified technical framework designed specifically to address those gaps.
- Methodological soundness: Leveraging TabPFN v2 for feature importance is a sophisticated choice, as it effectively incorporates the specific inductive biases required for tabular data into the pre-training phase.
- Comprehensive benchmarking: The evaluation is thorough, providing a detailed comparison against both state-of-the-art methods and established baselines across multiple datasets.
- Exhaustive ablation studies: The ablation experiments are particularly rigorous; by testing alternative discretization methods and kernels, the authors provide strong empirical justification for selecting the triangle kernel.
- Clarity and presentation: The manuscript is highly organized and technically detailed, supported by clear visualizations that make the complex multimodal architecture accessible.
- 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 methodological novelty appears incremental relative to the TIP architecture. To better showcase the significance of these enhancements, the authors should consolidate their results by including robust statistical reporting, such as standard deviations, confidence intervals, or p-values. Given that a 5-fold cross-validation was conducted, providing these metrics is essential to confirm that the observed 2%–4% gains are statistically significant and not due to stochastic variation.
Methodological clarifications:
- Aggregation formalism: Section 2.2 describes extracting importance from query, key, and self-attention matrices across 12 layers but lacks a formal definition of the “aggregation” process. The authors should provide the mathematical notation detailing how multi-head, multi-layer attention weights are collapsed into a final importance score (e.g., via Attention Rollout, simple averaging, or a specific weighted heuristic).
- Sample heterogeneity: Since TabPFN generates sample-specific attention maps, the authors should clarify how inter-row variability is managed. If the importance scores are averaged into a global metric, a justification is needed for bypassing inter-sample heterogeneity, which is often a critical factor in diverse medical cohorts.
- Loss formulation: The notation in Equation 2 is somewhat ambiguous; specifically, using “CE” and “KL” for a single sample-category item is misleading. A more rigorous formulation of the Discretized Reconstruction loss is required to improve technical clarity.
Suggestions for strengthening the contribution:
- Foundation model integration: While TabPFN v2 is utilized for feature importance extraction, the authors might consider whether using it directly as a backbone feature extractor—providing embeddings as model inputs in addition to masking weights—could further enhance performance.
Minor improvements:
- Visual consistency: If the feature importance is a primary driver for the masking mechanism, this should be explicitly illustrated or mentioned in the second diagram of Figure 1, where the masking process is currently omitted.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper provides a well-motivated extension of the TIP architecture. While the inclusion of additional statistical reporting (such as confidence intervals) would further consolidate the results, the proposed enhancements are clearly justified and meticulously detailed. The methodology is both intuitive and scientifically grounded; specifically, the strategic use of TabPFN’s inductive biases for feature importance and the implementation of a kernel-based soft-label discretization offer a convincing solution to the limitations of previous models. Furthermore, the thorough qualitative analysis and exhaustive ablation studies effectively strengthen the authors’ claims, demonstrating the clear value and practical utility of their approach in multimodal medical contexts.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper studies image-tabular pre-training for medical prediction tasks. The main idea is to make the tabular branch more semantically aware. Concretely, the method has two parts: (1) an importance-aware adaptive masking strategy to model inter-feature hierarchy with different diagnostic importance; (2) a soft-label discretization module to handle the continuity-discreteness duality within tabular features while preserving ordinal information.
Experiments are conducted on dermatology and ophthalmology datasets, with clear gains over the compared 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.
W1: The paper addresses image-tabular multimodal learning, which is an important direction in medical AI. W2: The method design is reasonable, and the motivation behind the two modules is easy to understand. W3: The empirical improvements are consistent across datasets. W4: The paper is generally well written and easy to follow.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
W1: The choice of frozen PFN is not fully justified. In particular, there is not an ablation that directly compares against using attention from a jointly trained ViT backbone. This comparison would help clarify whether freezing PFN is really needed.
W2: The behavior of masking score (s) during training is unclear. A visualization of how (s) evolves would make it easier to verify whether the method is actually attending to clinically relevant regions/features.
W3: The related-work discussion on ordinal information can be strengthened. The proposed soft-label discretization is intended to preserve ordinality, but prior work (e.g., [R1]) has similar motivation. A clearer positioning against this line of work would improve the paper.
W4: There is a technical description issue in Table 1: DINOv3 is self-supervised rather than supervised.
[R1] Yan, Jiahuan, et al. Making Pre-trained Language Models Great on Tabular Prediction. ICLR 2024.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I find the research question meaningful and the proposed method interesting. The paper addresses important challenges in image-tabular learning. However, the current empirical evidence is not yet fully convincing in several key aspects (e.g., the frozen PFN design choice and the training dynamics of the masking score). Additional ablation studies and deeper analysis would strengthen the claims and improve the overall quality of the work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for appreciating AID’s clinical importance and methodological grounding. The central contribution is a semantic-aware pre-training formulation that treats structured clinical variables as supervision signals for image-tabular medical representation learning. We address concerns below.
Q1: Novelty beyond TIP and relation to TP-BERTa (R2, R3). In contrast to TIP/MMCL/CITab’s semantic-agnostic flat-vector treatment with uniform masking and continuous regression, AID models inter-feature diagnostic hierarchy via a label-free importance curriculum and intra-feature ordinal-continuous structure via soft-label discretized reconstruction. We thank R3 for pointing to TP-BERTa, whose goal is orthogonal: it encodes ordinality at the input-tokenization level for a unimodal tabular LM, while AID enforces it at the reconstruction-objective level inside image-tabular multimodal pre-training, where visual evidence reconstructs masked clinical semantics. Empirically (Tab.1), AID’s linear probing exceeds prior SSL baselines’ full fine-tuning (incl. TIP, 0.984 vs 0.971), with OOD 0.944 vs 0.911 showing gains from semantic-aware design, not scaling. TP-BERTa will be cited and contrasted in camera-ready.
Q2: TabPFN role, aggregation, masking stability (R1, R2, R3). TabPFN serves only as an offline label-free meta-prior for dataset-level feature hierarchy. Designed for small-scale tabular data, it is applied once offline rather than as a backbone for 400K cohorts. For aggregation (R2), each layer averages multi-head weights; feature saliency from query (row sum, outgoing influence), key (column sum, incoming relevance) and self-attention (diagonal, self-relevance) is summed across layers, L1-normalized and min-max scaled, yielding a closed-form heuristic, not Attention Rollout. Formal notation will be added to §2.2.For row heterogeneity (R2), TabPFN’s in-context inference uses a single PCA-pseudo-label across rows, producing a dataset-level vector by construction. By design, it serves as a cohort-level masking prior, not a patient-specific explanation; patient-level heterogeneity is modeled by the multimodal encoder during representation learning. We clarify that s_i is fixed by design as the label-free curriculum; the representation evolves under it (Fig.3(e,f)). The per-feature cap α=r_max=0.7 (R1) applies only to top-importance features; the dataset-level expected ratio stays ~0.3, so adaptive masking targets clinically salient features rather than masking more. On joint training (R3), a jointly trained visual prior captures image-token saliency from multimodal optimization, whereas TabPFN supplies a tabular bias before pre-training (future work).
Q3: Soft-label discretized reconstruction (R1, R3). On bin sensitivity (R1), B∈{20,50,100} shows an inverted-U with B=50 consistently best (smaller B over-smooths intervals; larger B approaches noisy regression). Unlike hard quantile discretization, the triangle-kernel preserves ordinal neighborhood structure; unlike Gaussian smoothing it avoids over-smoothing while distinguishing near-bin from far-bin errors. The DR loss notation in Eq.2 will be refined to clarify summation over samples, features and bins (R2).
Q4: Statistical reliability and three-head protocol (R1, R2). Due to space, fold-wise std was omitted; in our 5-fold protocol AID’s std (≈0.5%) stays below TIP’s (≈1%), placing the OOD margin well outside fold-to-fold variation. The three-head ensemble averages image-only, tabular-only and multimodal logits, applied uniformly to AID and baselines (per TIP/CITab), giving no AID-specific advantage; representation quality is evidenced by linear probing surpassing prior baselines’ full fine-tuning.
Minor. DINOv3 is self-supervised in pre-training but grouped with supervised methods because we fine-tuned its weights with labels; we will relabel it as “SSL”. More implementation details and full results will be made publicly available.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
All reviewers have reached a consensus to accept the paper.
