Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Thyroid ultrasound (US) automation couples two competing requirements: global, geometry-driven reasoning for nodule delineation and local, texture-driven reasoning for malignancy risk assessment. Under cross-center domain shift, these cues degrade asymmetrically, yet most multi-task pipelines rely on a single shared backbone, often inducing negative transfer. In this paper, we characterize this interference across CNN (ResNet34) and medical ViT (MedSAM) backbones, and observe a consistent trend: ViTs transfer geometric priors that benefit segmentation, whereas CNNs more reliably preserve texture cues for malignancy discrimination under strong shift and artifacts. Motivated by this failure mode, we propose a lightweight family of decoder-side adapters, the Multi-Kernel Gated Adapter (MKGA) and a residual variant (ResMKGA), which refine multi-scale skip features using complementary receptive fields and apply semantic, context-conditioned gating to suppress artifact-prone content before fusion. Across two US benchmarks, the proposed adapters improve cross-center robustness: they boost out-of-domain segmentation and, in the CNN setting, yield clear gains in clinical TI-RADS diagnostic accuracy compared to standard multi-task baselines. Source code: https://github.com/MaziarSabouri/MKGA.git

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4058_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/MaziarSabouri/MKGA

Link to the Dataset(s)

ThyroidXL: https://papers.miccai.org/miccai-2025/paper/2024_paper.pdf DDTI: https://www.kaggle.com/datasets/dasmehdixtr/ddti-thyroid-ultrasound-images

BibTex

@InProceedings{SabMaz_MultiKernel_MICCAI2026,
        author = { Sabouri, Maziar AND Bayasi, Nourhan AND Rahmim, Arman},
        title = { { Multi-Kernel Gated Decoder Adapters for Robust Multi-Task Thyroid Ultrasound under Cross-Center Shift } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a novel architectural approach for mitigating negative transfer in multi-task medical image learning under domain shift, with a particular focus on the interaction between geometry-driven and texture-driven objectives. While prior work in multi-task learning largely relies on shared encoders and optimization-based conflict mitigation (e. g. , gradient surgery), this work identifies a major limitation, such strategies do not explicitly control how heterogeneous features are fused during decoding, where task interference is ultimately realized. The central contribution of this paper is a new class of decoder-side adaptation modules that explicitly regulate multi-scale feature integration in a task-aware and context-conditioned manner. The proposed method, Multi-Kernel Gated Adapter (MKGA), deviates from conventional skip-connection fusion by introducing a structured, three-stage transformation of encoder features before they are merged into the decoder. First, MKGA performs multi-kernel feature refinement, where parallel convolutions with different receptive fields are applied to each skip tensor to capture complementary spatial statistics. This design enables simultaneous modeling of fine-grained, high-frequency texture cues and broader contextual structures within a unified representation, addressing the inherent scale mismatch between classification and segmentation objectives. Second, the method introduces a context-conditioned gating mechanism that dynamically modulates skip features based on higher-level semantic information from deeper decoder states. Unlike static attention or channel reweighting schemes, this gating formulation explicitly conditions on cross-scale context, allowing the network to suppress artifact-sensitive or task-conflicting activations prior to fusion. This is particularly critical under domain shift, where shallow features often encode spurious correlations (e. g. , scanner-specific textures or overlay artifacts). By integrating semantic guidance into the gating process, MKGA enforces a form of feature consistency aligned with downstream objectives. Third, MKGA employs residual fusion to integrate gated skip features with decoder representations in a stable and parameter-efficient manner. This design avoids the instability of direct concatenation while preserving representational flexibility. Building on this, the paper further proposes ResMKGA, a residual bottleneck variant that applies channel-wise recalibration to the deepest encoder features prior to decoding. This extension specifically targets the locus of multi-task conflict in the shared latent space, providing an additional degree of adaptation without modifying the backbone architecture. An important technical novelty lies in the localization of task adaptation to the decoder, rather than the encoder or loss function. This contrasts with dominant paradigms such as backbone specialization, task-specific branches, or gradient-level conflict resolution. By operating directly on skip pathways, the proposed method leverages the hierarchical structure of encoder–decoder networks to introduce fine-grained, scale-aware control over feature propagation. Importantly, this design is orthogonal to backbone choice and optimization strategy, making it compatible with both convolutional and transformer-based encoders, as well as complementary techniques such as parameter-efficient fine-tuning (e. g. , LoRA) and gradient surgery. In addition to the architectural contribution, the paper provides a systematic empirical analysis that reveals a previously underexplored phenomenon: asymmetric task degradation under domain shift. Specifically, the authors show that transformer-based encoders preferentially preserve global geometric priors, benefiting segmentation, whereas convolutional architectures better retain local texture statistics, which are critical for classification. This observation not only motivates the proposed design but also contributes a new perspective on backbone-task alignment in multi-task medical imaging. Extensive experiments demonstrate that the proposed adapters yield consistent improvements in cross-domain generalization, outperforming both standard multi-task baselines and optimization-based conflict mitigation methods. Notably, the results show that architectural control of feature fusion can reduce reliance on gradient-level interventions, suggesting a shift toward structure-aware multi-task design as a more robust solution to negative transfer.

    Thus, the technical contributions of this work are: (1) the introduction of MKGA, a novel multi-kernel, context-gated decoder adapter for task-aware feature refinement; (2) the extension to ResMKGA, incorporating residual bottleneck adaptation for stabilizing shared representations, and (3) the conceptual advancement of decoder-centric multi-task learning, demonstrating that targeted control of feature fusion provides an effective and generalizable mechanism for mitigating task interference under domain shift. These contributions establish a new direction for designing robust, parameter-efficient multi-task architectures in medical image analysis.

  • 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). Novel decoder-centric design: Introduces MKGA, shifting multi-task conflict handling from encoder/optimization to decoder-level feature fusion, which is relatively unexplored and technically impactful. 2). Multi-kernel refinement: Uses parallel receptive fields to jointly capture global structure (segmentation) and local texture (classification), addressing task heterogeneity in a principled way. 3). Context-conditioned gating: Proposes a semantic-guided gating mechanism that suppresses artifact-prone skip features under domain shift-key for robustness. 4). ResMKGA extension: Adds residual bottleneck adaptation with channel recalibration, improving stability of shared latent representations with minimal overhead. 5). Insightful analysis of task interference: Demonstrates asymmetric degradation of geometry vs. texture tasks and differing CNN vs. ViT behavior-clear conceptual contribution. Strong cross-domain evaluation: Validated on in-domain + external datasets, directly addressing real-world generalization. Multi-task + clinical relevance: Jointly tackles segmentation, TI-RADS classification, and positioning, aligning with real diagnostic workflows. 6). Outperforms optimization-only methods: Shows architectural design (MKGA) is more effective than PCGrad, highlighting structural advantages. 7). Lightweight and modular: Parameter-efficient, easy to integrate into existing models and compatible with different backbones.

  • 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). Positioning relative to prior attention/gating modules could be clearer: While the proposed MKGA shares high-level similarities with prior designs such as Attention U-Net and Squeeze-and-Excitation Networks, the paper could more explicitly clarify the distinct role of decoder-level, context-conditioned gating under domain shift, and how this differs from existing skip-attention or channel recalibration mechanisms. 2). Perceived incremental nature of architectural components Some components (e. g. , residual refinement, channel recalibration) build on established design patterns. Emphasizing the novel combination and its targeted application to multi-task interference under domain shift would strengthen the perceived originality. 3). Limited comparison with broader domain generalization methods: The evaluation focuses on strong multi-task baselines and PCGrad, but could be expanded to include representative domain generalization approaches such as IRM or DomainBed. Including or discussing such methods would better contextualize the robustness claims. 4). Scope of validation limited to a single clinical application The experiments are focused on thyroid ultrasound. While this is a relevant and challenging setting, additional discussion (or future work) on generalization to other organs or modalities would strengthen the broader impact. 5). Clinical validation is limited to retrospective evaluation The study demonstrates improvements on clinically meaningful tasks (e. g. , TI-RADS), but does not include prospective validation or reader studies. Clarifying this as future work would align expectations with MICCAI’s translational goals. 6). Ablation studies could further disentangle component contributions: While ablations are provided, additional analysis on component interactions and task-specific effects could further strengthen understanding of why the method works.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses an important and practically relevant problem in medical image analysis, robust multi-task learning under cross-center domain shift. This is a well-motivated setting, and the paper clearly articulates the challenge of conflicting geometry-driven (segmentation) and texture-driven (classification) objectives. The empirical analysis highlighting asymmetric task degradation and differing behavior of CNNs versus transformers is insightful and adds conceptual value beyond the proposed method. From a technical perspective, the paper proposes a decoder-centric solution (MKGA/ResMKGA) that introduces multi-kernel refinement and context-conditioned gating for skip feature fusion. This is a well-designed and intuitive approach that targets an underexplored point of intervention (the decoder), and the method is lightweight, modular, and easy to integrate into existing architectures. The inclusion of a residual bottleneck variant further strengthens the design by addressing deeper feature conflicts. The experimental evaluation is a strong aspect of the paper. The use of both in-domain and external cross-center datasets, along with multiple tasks (segmentation, malignancy classification, and positioning), provides a comprehensive assessment. The results demonstrate consistent improvements in cross-domain robustness, and the inclusion of statistical testing and ablation studies adds credibility to the findings. However, the overall novelty is somewhat limited, as several components build on existing ideas such as attention gating, multi-scale convolutions, and channel recalibration. The paper would benefit from clearer positioning relative to prior work and stronger comparisons with domain generalization or advanced multi-task learning methods beyond PCGrad. Additionally, the validation is limited to a single application domain (thyroid ultrasound), and clinical validation remains retrospective. Overall, the paper is technically sound, well-executed, and addresses a meaningful problem, with strengths slightly outweighing its limitations. This justifies a borderline accept recommendation, contingent on improved positioning and clarification in rebuttal.

  • 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.

    Final recommendation is to accept.



Review #2

  • Please describe the contribution of the paper

    The paper addresses the “negative transfer” issue in multi-task learning for Thyroid ultrasound under cross-center domain shifts. The authors argue that segmentation and malignancy risk assessment require different feature properties that conflict during domain shift. To mitigate this, they propose MKGA and ResMKGA at the decoder skip connections to refine multi-scale features and suppress artifacts via semantic gating. Evaluations are performed on ResNet34 and MedSAM backbones using ThyroidXL and DDTI datasets

  • 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.Insightful Problem Motivation: The characterization of how CNNs and ViTs degrade differently under domain shift is well-motivated for the medical imaging community. 2.Lightweight Design: The proposed adapters are parameter-efficient, making them suitable for clinical deployment or fine-tuning foundation models. 3.Extensive Ablations: The authors provide a detailed breakdown of the components, which helps in understanding the source of performance gains.

  • 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.Limited Innovation: The MKGA architecture is incremental, resembling existing Attention U-Net and Inception-style fusion. It lacks fundamental theoretical breakthroughs, acting more as an ensemble of known modules than a novel paradigm. 2.Obsolete Baselines: ResNet34 is an outdated backbone that no longer represents SOTA. The absence of comparisons with modern models undermines the generalizability of the conclusions. 3.Poor Validation: External results on DDTI are underwhelming. ViT-based classification AUCs (0.4-0.5) suggest model failure, and the marginal gain in CNN AUC (0.577 to 0.62) lacks clinical significance, failing to support the “Robustness” claim. 4.Shallow Analysis: Attributing ViT’s cross-domain collapse solely to “weak texture” is insufficient. The paper fails to provide a deep analysis of why foundation models like MedSAM cannot learn robust features despite fine-tuning. 5.Unproven Artifact Suppression: Claims that the gating mechanism suppresses US artifacts lack empirical proof. No quantitative visualization or sensitivity analysis is provided to confirm that artifacts are accurately filtered.

  • 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.

    (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?

    While the paper addresses a relevant problem, its core contribution is hindered by a lack of substantial innovation. The proposed MKGA adapter represents a largely incremental improvement over existing architectures. More importantly: 1.Substandard Clinical Performance: Malignancy classification AUCs in cross-center testing are consistently low (0.5-0.6), indicating that the proposed method does not effectively resolve cross-center robustness issues in a clinically meaningful way. 2.Missing SOTA Comparisons: The study lacks comprehensive comparisons with mainstream Domain Generalization or advanced Multi-Task Learning balancing strategies.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal convincingly addresses my major concerns. I therefore revise my decision from Weak Reject to Accept. The justifications are as follows: Artifact suppression evidence: The NoGate ablation (Table 3) shows that removing the gate causes TI‑RADS accuracy on DDTI to drop from 63.2% to 49.9% while segmentation remains stable. This provides causal proof that the gate selectively suppresses artifact‑driven features. Clinical significance: Baseline ResNet34 achieves only 40.6% accuracy on DDTI, MKGA raises this to 63.2%, a meaningful recovery under heavy cross‑center shift. Novelty and baselines: The authors clarify MKGA’s distinction from SE and Attention U‑Net. ResNet34 is chosen for inductive bias analysis, not as a SOTA claim. The lack of modern CNN comparisons is a limitation but not fatal. The remaining concerns (modest absolute AUC, missing backbones) do not outweigh the rebuttal’s strength. The paper is now marginally above the acceptance threshold.



Review #3

  • Please describe the contribution of the paper

    Noticing how certain backbones are better at capturing geometry information for segmentation while others are better at texture detection for malignacy assessment, the authors argue that training a multitask learning framework with a single backbone can degrade performance. The authors introduced a family of decoder side adapters, MKGA and ResMKGA, to tackle these issues by gating and refining multi-scale skip features using semantic context to mitigate artifact-driven noise from shallow features.

  • 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.
    • good presentation of implementation details with promised code and model release
    • method is designed with the high artifact and noise nature of ultrasounds in mind, ablation rigorously justifies design
    • statistics using Wilcoxon, McNemar, DeLong, and FDR correction in comparison with baselines are extensive
  • 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.
    • As a cross-center framework, the authors demonstrated one in-domain dataset and one out-of-distribution test dataset. Understanding the limitation of data availability, it may be more beneficial to include more cross-center datasets to fully demonstrate the capacity of this method.
    • Given the objective of the paper, it would be ideal to include stronger baseline methodology specifically for cross center generalization and multi-task training, rather than merely comparing against backbone + PCGrad and LoRA.
    • The best segmentation and best diagnosis results do not come from one consistently superior configuration. Specifically, table 1’s DDTI results show a task tradeoff with MedSAM+ResMKGA+LoRA giving the best external segmentation Dice (0.675), while ResNet34+MKGA giving better external TI-RADS classification (AUC 0.642, accuracy 0.632).
    • Organization-wise, the experiments section introduces various important training settings which can use more thorough explanations (for instance, the frozen and unfrozen adaptability is a crucial distinction).
  • 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?

    The authors do an excellent job exploring a crucial problem of backbone inductive biases, especially in the context of geometry vs texture learning for noisy, high-artifact ultrasound imaging. The authors sets up extensive experiments to demonstrate the effectiveness of proposed lightweight decoder adaptation methods. However, the robustness claim will be much stronger with more cross-center data and stronger baselines with generalization and multitask training methods.

  • 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.

    Authors answered the concerns raised and makes good cases.



Author Feedback

We thank the AC & Reviewers for their insightful comments. 1.Novelty (R1, R2, AC) We respectfully argue that MKGA is not a recombination of prior modules; each component addresses a distinct gap: 1) Squeeze-and-Excitation performs global channel recalibration, leaving it blind to localized spatial artifacts like US calipers. MKGA’s gate is a full H×W map enabling localized suppression; 2) Multi-kernel refinement is deliberately minimal (two branches), calibrated to the asymmetric scale demands of segmentation vs. texture classification; 3) Standard Attention U-Nets use shallow features to guide deep ones. MKGA inverts this: deep semantic context filters artifact-corrupted shallow features before fusion. The combination targets the decoder-level locus of multi-task negative transfer, which is a novel, testable formulation. 2.Artifact suppression (R2, AC) Our gate isn’t a rigid binary mask (prone to overfitting to specific fonts), but a soft statistical spatially-localized recalibrator (~0.5 weights). This ‘dimmer switch’ uniformly dampens high-frequency artifact spikes from shallow features, preventing artifact-driven gradients from poisoning the shared latent space. Table 3 provides causal evidence. Removing the gate (NoGate) collapses TI-RADS on artifact-heavy DDTI (AUC: 0.589 vs. 0.642; Acc: 0.499 vs. 0.632, p<0.001) while segmentation stays stable (Dice: 0.673 vs. 0.659). This dissociation directly proves selective artifact suppression. Removing multi-kernel refinement (NoMulti) independently degrades both tasks (Dice: 0.629, Acc: 0.359, p<0.001), confirming both components are necessary. 3.Baselines: DG, MTL, and Backbone (R1, R2, R3, AC) Per MICCAI guidelines prohibiting new experiments, Tables 1 & 2 already benchmark against advanced baselines: 1) SOTA Backbone: MedSAM+LoRA (ranks 4, 16, 32) still fails on DDTI (AUC 0.455-0.528 vs. our 0.642); 2) Advanced MTL: PCGrad (Table 2) is the premier gradient-surgery baseline. ResNet34+MKGA (AUC 0.642, Acc 0.632) strictly outperforms ResNet34+PCGrad (AUC 0.633, Acc 0.594) on the same backbone, proving structural adaptation surpasses gradient surgery; 3) IRM/DomainBed require multiple source domains, inapplicable to our single-center setup (will clarify this explicitly). While ResNet34 enables CNN-vs-ViT inductive bias analysis, our backbone-agnostic MKGA empowers this lightweight model to drastically outperform the massive MedSAM. 4.Clinical Robustness and Task Trade-offs (R2, R3, AC) Robustness extends beyond AUC. Domain shift causes a severe calibration collapse; the baseline ResNet34 (unfrozen) drops to 40.6% Acc on DDTI (clinical unusability). MKGA rescues this to 63.2% Acc while preserving boundaries (Dice 0.659), a +22.6 pp absolute gain (McNemar p<0.001) of clear clinical significance. On R3’s trade-off and R2’s request for deeper ViT analysis: MedSAM achieves marginally better segmentation (Dice 0.675 vs. 0.659, p>0.05, non-significant) but collapses diagnostically (AUC 0.478, Acc 0.463; near random). ViT self-attention acts as a spatial low-pass filter. It aids coarse geometry but intrinsically smooths over high-frequency micro-textures (e.g., microcalcifications) essential for TI-RADS. LoRA cannot rewrite this core inductive bias, causing diagnostic failure under shift. A system with a non-functional diagnostic head has no clinical value. ResNet34+MKGA is the only configuration with simultaneously strong segmentation and viable TI-RADS classification. This trade-off validates, not undermines, our central hypothesis on backbone-task alignment. 5.Analysis Gap (Frozen/Unfrozen) & Clinical Scope (R1, R3, AC) Sec 3.4 will clarify ‘Frozen’ tests MKGA as a PEFT module on locked backbones, while ‘Unfrozen’ tests end-to-end synergy. Both are clinically meaningful, and their comparison isolates MKGA’s independent contribution. Scope: Artifact-driven conflict generalizes to echocardiography, breast, and musculoskeletal US. We will frame prospective validation as future work.



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.

    This paper proposes a decoder-centric adapter framework for multi-task thyroid ultrasound analysis under cross-center domain shift. The work focuses on mitigating negative transfer between segmentation and classification via task-aware feature fusion at the decoder level. The reviewers present mixed opinions. Reviewer #1 and #3 highlight the well-motivated problem, solid design, and strong experimental analysis, while Reviewer #2 is more critical, questioning the level of innovation, experimental validity, and robustness claims. Overall, the paper lies in a borderline range, with both supportive and critical viewpoints.

    The main weaknesses include:

    • Moderate novelty, with components building on existing attention, fusion, and gating mechanisms;
    • Limited comparison with broader baselines, particularly domain generalization and advanced multi-task methods;
    • Concerns about robustness claims, given relatively modest performance gains and limited external validation;
    • Some analysis gaps, including insufficient justification of certain design choices and limited empirical evidence for key claims (e.g., artifact suppression).

    Given the balance between solid experimental work and concerns about novelty and evaluation, I consider this paper borderline. The final decision should depend on how well the authors address the key concerns. Thus Invite Rebuttal.

  • 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 rebuttal has adequately addressed the main concerns from all reviewers and is sufficient to support acceptance. Therefore, I recommend Accept.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Congratulations!



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The reviewers are satisfied with the rebuttal.



back to top