List of Papers Browse by Subject Areas Author List
Abstract
Most mammography AI systems predict benign versus malignant from the image alone, ignoring the structured reasoning radiologists perform through the BI-RADS descriptor lexicon. We introduce a conditional modulation framework that dynamically shifts intermediate feature statistics based on BI-RADS descriptors. Rather than considering descriptors as static labels, our method produces \emph{sample-specific} modulation parameters: the same attribute generates different feature-space transformations depending on the image, capturing context-dependent diagnostic meaning. We show that our approach is highly effective. Across CBIS-DDSM and CMMD, our model improves AUC by up to +8.3\% and F1 by up to +5.0\% over the image-only baseline. By representing descriptors as conditional continuous embeddings, our framework enables exact interpolation between attributes within the same BI-RADS category; interpolation experiments confirm monotonic risk paths aligned with clinical severity, despite training only on hard categorical labels. Our model enables a post-detection decision-support mode using probabilistic outputs from emerging BI-RADS prediction models. Perturbation analysis confirms approximately 25\% of joint-setting predictions change when all descriptors are removed. \footnote{The code is available on \url{https://github.com/shahrma/semantic-mammo}}
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1039_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MahSha_Semantic_MICCAI2026,
author = { Mahpod, Shahar AND Ben Artzi, Gil},
title = { { Semantic Feature Modulation for Mammographic Lesion Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a conditional modulation framework that integrates BI-RADS descriptors into a ConvNeXt-Base backbone for mammographic lesion classification. Instead of treating descriptors as static categorical inputs (via concatenation or multi-task prediction), the method replaces the affine parameters of selected Layer Normalization layers with parameters predicted by descriptor-specific MLP modulation networks. The key insight is that BI-RADS descriptors are context-dependent: the same descriptor should produce different feature transformations depending on the image content. The paper also introduces continuous descriptor embeddings that enable sub-categorical interpolation and a second-opinion mode using probabilistic BI-RADS predictions.
- 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.
-Original approach: conditioning normalization layers on structured clinical descriptors is a well-motivated idea that has not been explored in mammography. The connection to FiLM-style conditioning is properly acknowledged, but the application to BI-RADS descriptors with hierarchical injection is novel. -Good experimental design: evaluation on two datasets (CBIS-DDSM and CMMD) with 5-fold CV over 5 independent runs provides robust statistical estimates. Standard deviations are reported, and the baselines (Concat, SE-Gate, MTL) represent the main alternative integration paradigms. -Clinically meaningful interpolation experiments: the monotonic risk trajectories from hard one-hot labels (mean score = 1.00 for all adjacent pairs) demonstrate emergent ordinal structure in the embedding space, which is a striking and useful finding. -Perturbation analysis is well done, confirming that approximately 25% of predictions change when descriptors are removed, and that lesion-specific descriptors (margins, shape) dominate, consistent with clinical knowledge. -Clear and well-organized writing with thorough mathematical formulation. The paper is easy to follow despite the multi-component architecture
- 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 operates in a post-detection setting with pre-extracted ROIs. While the authors acknowledge this, the practical impact of descriptor availability at inference time is not fully discussed. In real clinical workflows, not all descriptors are always available. -CMMD results are less convincing than CBIS-DDSM: on the mass classification setting, the Base model actually achieves the highest AUC (86.9%), outperforming both Cond-LN variants. This deserves more discussion. -No comparison with transformer-based backbones or more recent architectures. The exclusive use of ConvNeXt-Base limits generalizability claims. -The second-opinion mode using probabilistic descriptors is only conceptually described, not experimentally validated with actual automated BI-RADS prediction outputs
- 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?
I recommend acceptance due to the paper’s strong combination of methodological novelty, clinical relevance, and rigorous validation.
The proposed conditional modulation of normalization layers using BI-RADS descriptors is both original and well-motivated, moving beyond standard integration strategies. The experimental design is robust, and the results demonstrate not only performance gains but also clinically meaningful behavior, such as ordinal consistency and sensitivity to key descriptors. Overall, the work is technically sound, clearly presented, and offers a meaningful contribution to medical image analysis.
- 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 #2
- Please describe the contribution of the paper
The paper proposes a Semantic Feature Modulation framework to integrate BI-RADS descriptors into mammographic lesion classification. The core contribution is the use of Conditional Layer Normalization (Cond-LN), where BI-RADS descriptors are used to predict the affine parameters of Layer Normalization layers within a ConvNeXt backbone. The authors argue that this approach allows for context-dependent descriptors, meaning the same BI-RADS attribute (e.g., “spiculated margin”) can generate different feature-space transformations depending on the specific image content. Additionally, the framework claims to support exact sub-category interpolation, allowing the model to handle soft labels or probabilistic inputs representing uncertainty between BI-RADS categories (e.g., a mixture of “lobulated” and “irregular”), demonstrating monotonic risk paths aligned with clinical severity despite training on hard labels.
- 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 Integration of Clinical Knowledge. The paper presents a technically sound method to bridge the gap between image-based deep learning and the structured reasoning used by radiologists (BI-RADS lexicon). Unlike simple late fusion or multi-task learning, the proposed Conditional Layer Normalization allows the descriptors to dynamically modulate intermediate feature statistics. Support for Soft Inputs and Interpolation. The ability to handle convex combinations of BI-RADS attributes is a unique feature. The experimental validation showing monotonic risk increase during interpolation (e.g., from round to irregular shape) aligns well with clinical intuition and adds a layer of interpretability often missing in black-box models. The paper effectively adapts Conditional Layer Normalization (Cond-LN) to the medical domain. This is a strong technical approach for injecting structured prior knowledge (BI-RADS) into a deep learning backbone, offering a more nuanced alternative to simple feature concatenation. Handling Uncertainty: The proposed framework’s ability to support exact sub-category interpolation is a notable strength. By handling soft labels and probabilistic inputs, the method addresses the inherent inter-observer variability found in radiological reporting, which is a significant practical challenge in clinical deployment. Interpretability Potential: By modulating features based on specific descriptors, the method theoretically offers a path toward better interpretability, allowing researchers to visualize how specific semantic attributes (like “calcification type”) alter the feature space geometry.
- 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.
Clinical Feasibility & Circularity: The paper suffers from a potential “circularity” issue in clinical application. BI-RADS descriptors are typically the result of a diagnostic assessment, not a priori inputs. Unless the paper explicitly details a pipeline where these descriptors are predicted automatically or provided by a referring physician before the classification task, the method’s utility is limited to retrospective analysis rather than prospective screening. Besides, while the application is specific, the core mechanism—using side information to modulate normalization layers (Cond-LN/FiLM)—is a well-established technique in computer vision and generative modeling. The paper applies this existing formulation to mammography but does not introduce significant algorithmic innovations to the modulation mechanism itself. Also, the performance of the Semantic Feature Modulation is heavily dependent on the accuracy of the BI-RADS annotations. The paper does not sufficiently explore the model’s robustness to noisy or incorrect descriptor labels, which are common in real-world datasets. If the input descriptors are wrong, the feature modulation may actively degrade performance by suppressing relevant visual features.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
Due to concerns regarding the clinical practicality of the proposed framework and insufficient comparative analysis. While the application of Conditional Layer Normalization (Cond-LN) to modulate features based on BI-RADS descriptors is technically sound, the method relies on the availability of accurate BI-RADS annotations during inference. In real-world screening scenarios, these descriptors are often the output of the diagnostic process rather than known inputs; requiring them as prerequisites limits the method’s utility to retrospective analysis or specific decision-support sub-tasks rather than general screening. Furthermore, the experimental validation is not fully convincing. The performance gains over the baseline ConvNeXt model appear marginal, raising questions about whether the added complexity of the modulation module is justified. The paper also lacks a comparison against simpler feature concatenation or late-fusion baselines, which makes it difficult to ascertain if the specific mechanism of modulating Layer Normalization parameters is necessary or if standard fusion techniques would yield similar results. Finally, the claim of “exact sub-category interpolation” is theoretically interesting but lacks sufficient visualization or ablation to demonstrate its tangible impact on classification boundaries.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a conditional modulation framework that integrates BI-RADS descriptors into deep mammographic lesion classification in a context-dependent and continuous manner. Rather than treating BI-RADS descriptors (e.g., density, shape, margins) as static categorical labels, the paper adopts conditional Layer Normalisation (CLN). The sample-specific affine parameters (γ, β) of selected normalisation layers are derived from separate MLP projectors corresponding to the BI-RADS descriptors. The proposed model supports automated second-opinion interpretation using probabilistic outputs from emerging BI-RADS prediction models.
- 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.
In this paper, the strengths lie in its novel conditional normalisation formulation, the ability to interpolate within BI-RADS categories, the integration of probabilistic descriptors for second-opinion reasoning, strong empirical gains, and emergent clinical coherence in the learned embedding space. These contributions render the work both technically innovative and clinically well aligned. The major strengths are given as follows. 1.The paper proposes Conditional Layer Normalisation (CLN) guided by BI-RADS descriptors. Rather than using static concatenation or gating, the formulation replaces the affine parameters in Layer Normalisation with sample-specific modulation predicted from multiple BI-RADS descriptors. 2.By embedding descriptors as continuous vectors, the model supports convex combinations of attributes (e.g., 60% spiculated + 40% ill-defined). This represents a novel capability, as previous descriptor integration methods typically rely on hard categorical labels. Interpolation experiments demonstrate monotonic malignancy risk trajectories consistent with clinical severity, even when trained only on discrete labels. This emergent property is both technically impressive and clinically meaningful. 3.The proposed method can provide a reliable second-opinion interpretation. This allows radiologists to explore ‘what-if’ scenarios and uncertainty in descriptor assignment, making the system more transparent and interactive. Perturbation analysis shows that approximately 25% of predictions shift when descriptors are removed, demonstrating that the model meaningfully relies on descriptor conditioning. 4.The paper systematically evaluates the effect of descriptor ablation, revealing that mass margins and shape exert the strongest influence on predictions, which aligns well with established radiological knowledge. 5.Interpolation experiments show that the learned descriptor embeddings naturally rank malignancy risk in accordance with BI-RADS severity levels.
- 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.This is an incremental novelty in architecture, using conditional normalisation adapted to BI-RADS. The scope of the experimental dataset is limited, the number of cases in each dataset is very small, and to some extent, it cannot demonstrate the generalisability and reproducibility of the designed model. Additionally, reliance on various descriptors may hinder feasibility, and clinical validation is lacking. While technically sound and conceptually interesting, it has not yet demonstrated the level of clinical readiness or methodological breakthrough expected for top-tier acceptance. 2.Improvements over baselines are relatively small in absolute terms (AUC +2–4% in many settings). While statistically consistent, these gains may not be compelling enough to justify acceptance on the basis of novelty when compared to state-of-the-art multimodal fusion methods. 3.The paper uses Grad-CAM++ visualisations, which are standard and widely used. The adoption of more advanced interpretability methods (e.g., counterfactual explanations, prototype-based reasoning) would have provided stronger evidence of clinical trustworthiness. 4.The paper compares against concatenation, SE-Gate, and multi-task learning, but omits recent multimodal fusion approaches (e.g., Deep BI-RADS Network [Ben-Artzi et al., 2025], MedNeXt [Roy et al., MICCAI 2023]). It is necessary to provide comparative experiments with state-of-the-art (SOTA) methods. 5.Claims in the paper such as “trustable second-opinion mode” and “emergent clinical coherence” are strong but not supported by clinical trials or radiologist feedback. As such, the language oversells the contribution relative to the evidence provided. 6.The ablation experiments are insufficient. Why is the CLN set to 27 blocks? Please provide relevant validation experiments to explain this choice. 7.The paper presents a lesion classification model that integrates BI-RADS descriptors (such as density, mass shape and margin). However, in practice, the BI-RADS descriptor lexicon for lesions includes other features (e.g., morphology and distribution of calcification, bilateral breast structure and symmetry, etc.), and these descriptors have not been considered. Could you explain the reasons for this? 8.No radiologist-in-the-loop validation is provided. This is necessary to establish the clinical usability of the proposed model. 9.Due to the inherent limitations of mammographic imaging, in clinical practice, it is often necessary to compare CC and MLO views simultaneously to examine lesions and analyse their benign or malignant nature. However, the model only utilises a single lesion region combined with BI-RADS descriptors to adjust the benign/malignant classification. Is this reliable for clinical practice? 10.The model uses a unified input size of 384x384.Could this affect the detailed information of small lesions and lead to incorrect classification results? Please provide relevant experimental validation to address this. 11.In the deep learning era, large cohorts are generally required. However, this study uses only two datasets with uneven data splits, both of which are relatively small in scale. Please provide further experimental validation on unseen cross-domain external datasets.
- 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
Nothing.
- 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?
While the paper is technically innovative and clinically well aligned, with a novel descriptor-conditioning mechanism, strong empirical gains, and emergent semantic coherence, its novelty remains incremental in the context of existing conditional normalization literature. Furthermore, the dataset scope is limited; other relevant BI-RADS descriptors are not considered, its clinical significance is considerably overstated, its malignant classification performance is modest at best, and it lacks analysis of hard cases. Overall, the methodological contribution is only incremental. As such, the work is insufficient for direct acceptance and would require further experimental validation to address its identified weaknesses.
- 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
Author Feedback
We thank the Meta-Reviewer and Reviewers for their constructive feedback and for recognizing that our work offers “a novel perspective in this application domain”, is “well-motivated”, and is “experimentally solid”. We use this response to clarify the main intended use case and several points raised in the reviews.
Intended clinical setting. We agree that BI-RADS descriptors are produced during diagnostic interpretation and are not available as primary screening inputs. Our intended setting is therefore post-detection decision support: a second-reader tool that conditions malignancy assessment on radiologist-provided descriptors, including uncertain or partially specified ones. Two components of the submitted method support this use case: the learned “unknown” token for missing descriptors and the convex-combination descriptor interface described in Sec. 3.3.The interpolation analysis in Sec. 4.6 further shows that probabilistic descriptor inputs produce clinically coherent malignancy trends. Coupling the model to an automated upstream BI-RADS descriptor predictor is an important future direction, and we will clarify this deployment scenario in the final version.
Empirical gains and variability. The improvements over the image-only baseline are consistent in most settings but not uniform, as the reviewers note. Our Cond-LN-Inter is strictly highest AUC in 5/6 dataset×setting combinations compared with non-conditioning baselines, with AUC gains over Base ranging from +2.2 to +4.5 on CBIS-DDSM and from +4.3 to +8.3 on CMMD. The CMMD-mass case flagged by R1 (Base 86.9±0.3 vs. Cond-LN-Inter 86.3±0.6 over 5 runs) is a limitation we will sharpen in the final version: the difference is modest relative to the reported run-to-run variability, and this subset’s imbalanced malignant prior may further constrain how much headroom descriptor signal can add.
Contribution framing. We do not claim that conditional modulation is novel in general. Our contribution is the specific conditional design for BI-RADS: hierarchical injection across network depth according to descriptor semantics (global vs. lesion-level) and a soft-input interface supporting intra-category convex combinations of attributes. Together, these design choices and their specific realization enable the model to learn an emergent ordinal structure aligned with clinical reasoning (Sec. 4.6: monotonic risk paths despite training only on hard labels).
Additional reviewer-specific points. The concatenation baseline is reported in Tables 1–2, and our method outperforms it in the main CBIS-DDSM settings. Calcification descriptors are described in Sec. 4.2: density, morphology, and distribution descriptors; the joint setting uses all six plus lesion type. The use of 27 blocks follows the third stage of the ConvNeXt-Base architecture rather than being a tuned hyperparameter. On descriptor robustness, the submission evaluates related practical cases of descriptor uncertainty: adjacent-category ambiguity and probabilistic inputs (Sec. 4.6), missing inputs (Sec. 4.5), and unknown inputs (Sec. 3.3). Further robustness analysis would be valuable. Finally, “trustable second-opinion” will be softened to “second-reader candidate” in the final version.
We appreciate the suggestions raised by the reviewers regarding multi-view modeling, more baselines, and larger-scale validation. These are important next steps.
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.
This paper presents a well-motivated and technically sound methodological contribution that integrates structured clinical knowledge, incorporating Breast Imaging Reporting and Data System (BI-RADS) descriptors into a ConvNeXt network for deep mammographic lesion classification. Across the three reviews, there is consistent agreement that the work is clearly written, experimentally solid, and offers a novel perspective in this application domain. The strongest support comes from Reviewer #1, while Reviewers #2 and #3 provide weaker but still positive recommendations, leading to an overall leaning towards acceptance.
The main factors supporting this decision are the originality of adapting conditional normalisation to descriptor-driven modulation in a clinically meaningful way, and the ability to model context-dependent transformations rather than static feature fusion, for example concatenation. The continuous embedding of descriptors and the resulting interpolation behaviour are particularly compelling, as they demonstrate emergent ordinal structure aligned with clinical reasoning. In addition, the experimental design is reasonably thorough, including cross-validation, ablations, and perturbation studies that indicate the model is using descriptor information in a meaningful and clinically consistent manner.
There are, however, several concerns raised by the reviewers. A central issue is clinical feasibility, as the method assumes access to BI-RADS descriptors at inference time, which are typically outputs of the diagnostic process rather than inputs, as noted by Reviewer #2.The empirical gains, while consistent in some settings, are modest (Reviewers #2 and #3). There are also gaps in comparison with stronger or more recent baselines, limited validation of the proposed second-opinion setting with real probabilistic descriptor inputs, and insufficient analysis of robustness to noisy descriptors (Reviewer #2). Additional concerns include limited dataset scale, simplified single-view inputs, and missing justification for certain architectural design choices (Reviewer #1).
Therefore, I recommend Provisional Accept. In particular, the authors should address these issues, for example by explaining how the method would be deployed in practice, such as in conjunction with automated descriptor prediction or as a decision-support tool, and by providing deeper analysis of performance variability and limitations. Addressing these points will be important to confirm that the proposed approach offers practical value beyond its conceptual contribution.
