List of Papers Browse by Subject Areas Author List
Abstract
Ultrasound-based thyroid malignancy assessment requires both accurate nodule delineation and discriminative diagnosis. Joint segmentation and classification is therefore a promising solution to reduce error propagation and improve diagnostic consistency. However, existing joint pipelines often couple tasks mainly through shared parameters or prediction-level cues, providing limited feature-level interaction between branches. This weak coupling can lead to attention-boundary mismatch and poorly aligned cross-task representations. We propose IC-MTL, an interaction-aware contrastive multi-task framework that explicitly models bidirectional cross-task interaction and introduces supervised contrastive regularization on the post-interaction representations to jointly encourage lesion-aligned attention and class-separable diagnostic features. IC-MTL is evaluated on the public DDTI benchmark and a real-world clinical dataset. Across both datasets, IC-MTL improves performance over staged baselines and recent state-of-the-art joint models. Qualitative analyses further indicate improved lesion-aligned attention and more separable representations, supporting more reliable clinical risk stratification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6427_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{LinCha_InteractionAware_MICCAI2026,
author = { Lin, Chaochao AND Boumaraf, Said AND Liu, Xiabi AND Liu, Qianglin AND Niu, Lijuan AND Werghi, Naoufel},
title = { { Interaction-Aware Contrastive Multi-task Learning for Joint Segmentation and Malignancy Classification in Thyroid Ultrasound } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes IC-MTL, a multi-task framework for joint thyroid nodule segmentation and malignancy classification in ultrasound images. It introduces a bidirectional Cross-Task Interaction (CI) module that utilizes source-derived gating to transfer boundary information to the classification branch and malignancy semantics to the segmentation branch. It integrates a Supervised Contrastive Multi-Task Regularizer (SCMR) applied to post-interaction features at both pixel and image levels to align the representation geometry across 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.
The framework explicitly addresses the inherent clinical interdependence between anatomical delineation and pathological diagnosis. The CI module formalizes bidirectional message passing as a mathematically clear, gated residual update. Applying supervised contrastive objectives symmetrically on post-interaction features successfully shapes feature separability at multiple granularities.
- 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 CI gating formulation lacks algorithmic novelty and conceptually duplicates established multi-task paradigms like cross-stitch networks, sluice networks, and FiLM-like modulation. The evaluation completely omits mandatory quantitative comparisons against these classic feature-exchange architectures and recent thyroid-specific joint models. The methodology leaves critical technical details undefined, including backbone architecture, SCMR projection head dimensions, and pixel-level subsampling strategies for managing foreground/background imbalance. The dynamic loss reweighting scheme relies on an arbitrary Dice score threshold, introducing unfair training heuristics. The study lacks statistical significance testing and fails to report computational overhead, inference time, or memory consumption constraints.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 framework targets a highly relevant clinical bottleneck and demonstrates consistent empirical gains through logically motivated task coupling. However, the core algorithmic innovation remains marginal compared to existing cross-task feature modulation literature. The empirical validation is fundamentally compromised by the exclusion of direct comparisons with established multi-task baselines and thyroid-specific joint networks. The absence of statistical significance testing makes the modest performance gains unreliable. Missing architectural specifics and memory management details for the pixel-level contrastive loss severely hinder both reproducibility and real-world clinical deployability.
- Reviewer confidence
Somewhat confident (2)
- [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 proposes IC-MTL, an interaction-aware contrastive multi-task learning framework for joint thyroid nodule segmentation and malignancy classification in ultrasound images. The framework introduces two tightly coupled components: (1) a bidirectional Cross-task Interaction (CI) module that explicitly exchanges boundary-sensitive segmentation features into the classification branch (Seg→Cls) and malignancy-discriminative classification semantics into the segmentation branch (Cls→Seg) via a learned gated residual message-passing mechanism; and (2) a Supervised Contrastive Multi-task Regularization (SCMR) component that applies supervised contrastive objectives at both the pixel level (foreground/background separation for segmentation) and image level (benign/malignant separation for classification) on the post-interaction feature representations. The framework is evaluated on the public DDTI benchmark and an independent clinical dataset from a local hospital, demonstrating improvements in malignancy classification AUC and accuracy over staged and joint baselines while maintaining competitive segmentation DSC.
- 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.Principled bidirectional cross-task interaction with clearly defined roles for each direction. The central contribution of IC-MTL is the explicit formalization of information flow between the segmentation and classification branches in both directions. The design is grounded in a clear clinical rationale: boundary and shape cues from segmentation are informative for malignancy assessment (per TI-RADS morphological descriptors), and malignancy-related semantics can sharpen segmentation by refining attention to lesion-relevant regions. This bidirectionality is not merely architectural but is implemented with role-specific message-passing operators and validated through direction-specific ablation. Removing Seg→Cls reduces AUC by 5.53 percentage points (85.92 → 80.39), and removing Cls→Seg reduces DSC by 0.66 points (92.38 → 91.72), confirming that each direction contributes to its intended target task in the expected direction. 2.SCMR applies contrastive regularization at both task-appropriate granularities. The supervised contrastive regularization is applied at the image level for classification (pulling benign/malignant embeddings together within class and apart between classes) and at the pixel level for segmentation (pulling foreground/background pixel embeddings apart). This dual-granularity contrastive framework is conceptually consistent with the different feature geometries required by each task and is supported by the ablation: removing SCMR (Cls) drops ACC from 90.00 to 84.67, and removing SCMR (Seg) drops DSC from 92.38 to 90.72, suggesting that the two contrastive terms have complementary and largely non-overlapping effects. 3.Thorough ablation study with component-level granularity. Table 2 includes ten experimental conditions covering: classification-only baseline, staged pipeline baseline, full joint model without proposed components, removal of CI entirely, removal of each CI direction independently, removal of SCMR entirely, and removal of each SCMR term independently. This level of ablation granularity significantly strengthens confidence in the contribution of each proposed component and provides clear evidence for both CI and SCMR as necessary elements. 4.Rich qualitative analysis supporting mechanistic interpretation. Figure 3 provides t-SNE feature visualization comparing class separability across competing methods (Chowdary et al., Luo et al., He et al., and IC-MTL before and after CI), Grad-CAM attention maps for both DDTI and local datasets, and segmentation feature map comparisons before and after CI. Together, these analyses provide an intuitive and multi-faceted verification that CI improves feature separability, attention-boundary alignment, and lesion-focused segmentation activations in the expected ways, going beyond what tables alone can convey. 5.Evaluation on two datasets with different characteristics. Testing on both the public DDTI benchmark and an independent clinical dataset collected at a separate hospital provides a reasonable assessment of generalization under scanner and population shift. The two-dataset evaluation is particularly valuable because the local clinical dataset includes ultrasound video sequences rather than static images, adding a degree of real-world complexity. 6.Dynamic loss re-weighting strategy is practically sound. The two-stage training schedule that prioritizes segmentation until DSC ≥ 0.8, then shifts emphasis to classification, is a pragmatic solution to the common problem of task imbalance in multi-task learning. The rationale — stable spatial representation first, then discriminative semantic learning — is aligned with how clinicians sequentially assess thyroid nodules (first locate, then characterize), and it reduces the risk of classification branch overfitting on unreliable segmentation features early in training.
- 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.Both evaluation datasets are relatively small, limiting the generalizability of conclusions. The DDTI dataset contains 347 images from 299 patients, and the local clinical dataset contains 4,598 images from 116 patients. While five-fold patient-level cross-validation is used to prevent leakage, the effective validation set per fold is approximately 60–70 images for DDTI and approximately 23 patients for the local dataset. The AUC standard deviations of ±6.77 (DDTI) and ±6.95 (local) reflect this limited sample size and indicate that performance estimates are not stable across folds. For a malignancy classification system intended for clinical use, validation on 100+ independent test cases would be a more credible basis for clinical conclusions. Furthermore, neither dataset constitutes a prospective or multi-center cohort. 2.Hyperparameter sensitivity is not analyzed. The model involves several hyperparameters with non-trivial impact on performance: the DSC threshold (0.8) that triggers the loss re-weighting stage transition, the two sets of loss weights (α, β, γ, γ_seg, γ_cls) for each training phase, and the contrastive temperature τ = 0.5.The paper states these were “chosen empirically on the validation folds,” but no sensitivity analysis is provided. It is not clear whether the reported performance gains are robust to moderate perturbations of these values, or whether they depend critically on the specific configuration. A brief sensitivity sweep over at least the DSC threshold and τ would strengthen the reproducibility and robustness claims. 3.Performance on the local dataset shows non-negligible DSC degradation relative to the best baseline. On the local clinical dataset, IC-MTL achieves DSC 87.77%, whereas He et al. achieves DSC 88.28%. While the paper characterizes this difference as “marginal,” it is worth noting that the local dataset represents a more challenging real-world setting (acquired from video sequences, different scanner characteristics, independent clinical site), and that the segmentation branch in IC-MTL is jointly optimized with classification, which may introduce trade-offs. A more detailed analysis of the cases where IC-MTL’s segmentation falls behind — e.g., whether these are systematically the difficult malignant nodules with irregular boundaries — would inform the clinical interpretation. 4.No investigation of the interaction between CI and different encoder architectures. All experiments use the same shared encoder backbone. The generalizability of the CI mechanism across different encoder architectures (e.g., transformer-based vs. CNN-based encoders) is not assessed. As the field moves toward foundation model backbones for ultrasound, it would be valuable to know whether the gated residual CI operator remains effective when features are extracted from, e.g., a pretrained SAM or MedSAM encoder. 5.Code and data release is not planned. The paper does not mention any plan to release code or dataset. While the local clinical dataset cannot be released for privacy reasons, the code for IC-MTL should be made available for community use and reproducibility. The lack of a code release commitment limits the practical impact of this contribution. 6.The Grad-CAM visualizations in Figure 3b are presented on a small scale. The qualitative results in Figure 3b are presented at a resolution that makes it difficult to verify the claimed attention-boundary consistency improvements, particularly for the local dataset examples. Enlarged selected examples with clearer visual comparisons would substantially improve the interpretability of these qualitative results.
- 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 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
IC-MTL is a well-executed and clearly motivated framework. The main recommendation for improvement is to expand the evaluation to a larger cohort, ideally including at least one additional external site, to substantiate the generalization claims with greater statistical confidence. Publishing the code would also significantly increase the community uptake of this work. Adding hyperparameter sensitivity analysis — even as a brief supplementary table — would improve the reproducibility claim. For future work, exploring whether the CI module generalizes to stronger or pretrained encoder backbones (e.g., SAM-based) would be a valuable direction given the current momentum in foundation model adaptation for ultrasound analysis.
- 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?
IC-MTL presents a methodologically sound and clearly motivated framework for joint thyroid nodule segmentation and malignancy classification. The bidirectional cross-task interaction module addresses a genuine limitation of prior joint models — the asymmetric or implicit coupling between tasks — and the supervised contrastive regularization at dual granularities provides a principled approach to multi-task feature space alignment. The component-level ablation is thorough, the qualitative analyses are rich and informative, and the framework demonstrates consistent improvement over representative baselines on two datasets. The main limitations are the small size of both evaluation datasets, the absence of hyperparameter sensitivity analysis, and the lack of code release. These limitations reduce the confidence with which clinical generalization claims can be made, but do not undermine the methodological contribution. The paper is marginally above the acceptance threshold, and its acceptance would be strengthened by rebuttal clarifications regarding the local dataset size context and any plans for code release or expanded validation.
- 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 proposes bidirectional cross-task interaction (CI) and supervised contrastive multi-task regularization (SCMR) for a multi-task framework for joint segmentation and malignancy classification in thyroid ultrasound, thereby strengthening feature-level coupling between the two tasks and improving overall diagnostic performance.
- 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 problem setting is clinically well motivated, as nodule boundary and shape information are closely related to malignancy assessment in thyroid ultrasound, making joint segmentation and classification a reasonable and meaningful direction. 2.The proposed design is conceptually clear: rather than relying only on a shared encoder, it introduces bidirectional cross-task interaction (CI) to explicitly exchange features between the two branches, and SCMR to make the post-interaction representations more discriminative. 3.The paper includes fairly thorough and appropriate ablation studies, including removal of CI, directional interaction terms, and SCMR, which helps support the contribution of each component. 4.The method is evaluated on both a public benchmark (DDTI) and a local clinical dataset, which is a strength compared with studies validated on only a single dataset. 5.The paper provides qualitative analyses such as t-SNE, Grad-CAM, and feature-map comparisons, which help illustrate the intended effect of the proposed interaction mechanism beyond the reported quantitative metrics.
- 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 paper argues that CI improves classification by transferring segmentation-informed morphology, and Fig. 3b is presented as qualitative evidence of improved attention–boundary consistency. However, it is not obvious that the Grad-CAM visualizations clearly support this claim; in some cases, the heatmaps from He et al. appear at least as well aligned with the lesion mask. Could the authors provide stronger, preferably quantitative, evidence that the classification gains are specifically driven by morphology-informed interaction rather than generic multi-task regularization?
2.Although the paper seems to prioritize classification in practice, this is not clearly justified in the framing of the joint task. On DDTI, IC-MTL improves both DSC and classification metrics, whereas on the local dataset it improves classification but not segmentation over He et al. As a result, the claimed advantage of the joint framework is somewhat difficult to interpret. A stronger evaluation would be true cross-dataset external validation rather than separate within-dataset cross-validation.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not 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.
(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?
Please refer to the weaknesses above. If these concerns are adequately addressed, I would be open to raising my score.
- Reviewer confidence
Somewhat confident (2)
- [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
N/A
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 proposes IC-MTL, an interaction-aware multi-task framework for joint thyroid nodule segmentation and malignancy classification, leveraging bidirectional cross-task interaction and contrastive regularization. The reviewers generally agree that the paper addresses a relevant and meaningful problem and presents a well-structured multi-task framework. Reviewer #1 and #2 is clearly positive, highlighting strong ablation studies and consistent gains, while Reviewer #3 is more cautious, mainly due to concerns about novelty and missing comparisons. The overall assessment lies around the weak accept borderline, with some disagreement on the strength of the contribution.
The main strengths are:
- A clinically grounded formulation that explicitly models bidirectional interaction between segmentation and classification;
- A coherent and well-motivated framework, combining cross-task interaction and contrastive learning;
- Thorough ablation studies and qualitative analysis, supporting the contribution of individual components;
- Evaluation on both a public dataset and an independent clinical dataset.
The main weaknesses include:
- Limited novelty at the component level, with similarities to prior multi-task interaction mechanisms;
- Missing comparisons with some relevant multi-task baselines and limited analysis of generalization.
Despite some limitations in novelty and evaluation scale, the paper presents a well-executed and clinically meaningful multi-task framework, with consistent empirical improvements and solid experimental support. The contribution is considered sufficient for acceptance, particularly given its relevance and methodological clarity.