Abstract

Longitudinal CT imaging analysis, which captures the spatiotemporal evolu-tion of tumor characteristics across early treatment cycles, holds great promise for predicting immunotherapy response and optimizing personalized treatment strategies in advanced non-small cell lung cancer. However, existing methods that rely on shared-weight encoders often suffer from feature entanglement be-tween temporally stable and treatment-induced dynamic representations, where the dominant static features may suppress subtle dynamic signals that are most predictive of immunotherapy response. To address this challenge, we propose a Temporal-Calibrated Feature Disentanglement Network (TCFD-Net), which explicitly disentangles longitudinal CT features into temporal-invariant and temporal-variant components via a Temporal Invariant-Variant Disentangle-ment Module, supervised with cross-time mean squared error (MSE) con-sistency loss, batch distance loss and subspace orthogonality loss to ensure ef-fective disentanglement. Additionally, a Time-Gating Mechanism is designed to adaptively calibrate temporal-variant features based on inter-scan intervals, addressing the irregular scanning protocols common in clinical practice. Exper-iments on both internal and external testing dataset demonstrate that the pro-posed method out-performs existing multi-timepoint analysis approaches, achieving AUC improvements of 7.72%-18.84% and accuracy improvements of 10.72%-16.27% on the external testing dataset.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3685_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/YFSoaring/TCFD-Net

Link to the Dataset(s)

N/A

BibTex

@InProceedings{FuYao_TCFDNet_MICCAI2026,
        author = { Fu, Yao AND Gao, Ziwei AND Lu, Yinuo AND Xu, Weixin AND Qi, Mingxin AND Wu, Yukun AND Mu, Wei},
        title = { { TCFD-Net: Temporal-Calibrated Feature Disentanglement Network for Dual-Timepoint CT-Based Immunotherapy Response Prediction in Advanced Non-small Cell Lung Cancer } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This manuscript introduces a time-calibrated feature disentanglement network for characterizing changes between pre- and post-treatment CT scans for early prediction of immunotherapy response in advanced lung cancer.

  • 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.

    Strengths include: (1) The need to better encode temporal features, particularly when assessing treatment response, is well-motivated, and the general approach seems promising. (2) The inclusion of over 1,100 patients with advanced lung cancer who received immunotherapy across two independent centers. (3) t-SNE visualization provides compelling qualitative support that the proposed approach can disentangle treatment-induced dynamic and static information.

  • 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 evaluation was performed on a single split during internal validation and an external validation set. The lack of resampling means that only point estimates are given in Tables 1-3.It is unclear how reliable these estimates are without confidence intervals. The manuscript lacks an explicit discussion of the limitations.

  • 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

    Details on which temporal variants and invariant features are extracted by the two parallel 3D encoders are not specified and should be elaborated upon.

  • 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 need for improved methods to characterize temporal changes and predict treatment response is clearly established. The temporal disentanglement module is potentially innovative. Nevertheless, the explanation of temporal variant and invariant features is not clear, and the lack of resampling and limited reporting of metrics (just AUC and accuracy, not sensitivity/positive predictive value) reduces the completeness of the reported results.

  • 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.

    The authors provided adequate clarifications that address the main concerns related to performance reporting and clarity around the methods.



Review #2

  • Please describe the contribution of the paper

    This paper introduces TCFD-Net (Temporal-Calibrated Feature Disentanglement Network), a deep learning framework for predicting immunotherapy (IO) response in advanced lung cancer using longitudinal CT scans. The key contribution is novel, important, and biologically grounded – certain patient-specific features remain invariant across pre- and post-treatment scans and can dilute the subtle, treatment-induced changes most predictive of response. TCFD-Net explicitly disentangle temporal-variant (treatment-related) and temporal-invariant (patient-specific) representations through a novel architecture and loss functions. The architecture also incorporates a time-gating mechanism that calibrates features based on actual scan intervals, accounting for non-uniform imaging schedules. Experiments on internal and external validation cohorts demonstrate that TCFD-Net significantly outperforms existing methods.

  • 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.

    Novelty: The core idea of explicitly disentangling static and dynamic features is a clever and important contribution to the field. Beyond building predictive models, this approach has immense potential to elucidate biological underpinnings of time-invariant and treatment specific changes through multi-omic correlative analyses (eg. between time varying features and pathomic/genomic measurements).

    Strong Experimental Results: The performance improvements (AUC and accuracy) over existing methods are substantial and consistent across both internal and external datasets. This highlights the value of the proposed approach.

    Interpretability: The visualization of feature distributions using t-SNE demonstrates a clear separation of invariant and variant features providing insight into how the network is operating and validating the disentanglement process.

  • 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.

    Loss Function Weighting: The paper mentions using loss function weights (λ values) during training but doesn’t elaborate on how these weights were determined. This is a critical aspect of training and the lack of detail makes it difficult to reproduce the exact results or provide an idea of contribution of individual losses.

    Limited details on implementation: While the paper describes the general architecture, more details on the specific implementation of the 3D encoders, Swin Transformers, and Time-Gating Mechanism (units in the MLP) would be helpful for reproducibility.

    Clarity on time between CT scans: Typically, immediate post-treatment CTs are acquired every 6-9 weeks post-IO initiation. It would be helpful to include the median and range of time between pre and post-treatment CT scans in the dataset to understand the variance in scan intervals.

    Potential for overfitting: While demonstrating good external validation, the complexity and number of loss functions used in training increase the risk of overfitting, particularly since the training set was from a single institution and had modest size (534 patients). Was any data augmentation applied?

    Histology: Please use specific terminologies for if the patients had advanced “small cell” or “non-small cell” lung cancer, or whether the dataset included both histologies.

    Typo: Effectiveness and Interpretability of TIVDM are presented in Table 2, the text mentions table 1.

  • 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.

    (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 core idea of explicitly disentangling static and dynamic features from longitudinal CT scans is a clever and genuinely novel contribution to the field. This approach has immense potential to elucidate the biological underpinnings of time-invariant and treatment-specific changes through multi-omic correlative analyses, for example by linking temporal-variant features to pathomic or genomic measurements, and may aid in identifying mechanisms of immunotherapy resistance in non-responders. This approach can also be used to identify, for example, immunotherapy vs chemotherapy specific biological pathways with the right dataset, and its applications extend beyond predictive utility. Moreover, the performance improvements in AUC and accuracy over existing methods are consistent across both internal and external validation cohorts, highlighting the robustness and generalizability of the proposed approach.

  • 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.

    The authors have addressed all the comments and I’m happy to recommend acceptance of the paper.



Review #3

  • Please describe the contribution of the paper

    (1) A dual-phase CT prediction framework with decoupling of temporal variation characteristics was proposed; (2) A time calibration mechanism based on scan intervals was introduced; (3) Two-center validation was conducted, demonstrating favorable external test 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) Significant improvement in external testing performance; (2) Comprehensive method design with closed-loop logical chain.

  • 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) Figure 1 lacks aesthetic appeal and clarity in presentation; (2) The formulas and variables in the text fail to specify variable dimensions and domains, hindering formula comprehension; (3) The paper only reports AUC and ACC, which are insufficient for medical prediction tasks. Additional metrics such as SEN, SPE, 95% CI, and Delong test should be included; (4) Reference errors exist, e.g.: “Treatment responses were categorized per RECIST 1.1 criteria, grouping Complete Response (CR) and Partial Response (PR) as Responders, while Stable Disease (SD) and Progressive Disease (PD) were classified as Non-responders [19].” Reference [19] does not adhere to RECIST 1.1 criteria. (5) The clinical modeling is insufficient, lacking detailed definitions of key data parameters. For example: Is the CT input the entire thoracic volume, tumor patches, or ROI crops? Does the tumor require manual segmentation or automatic detection? How should patients with multiple lesions be handled? These issues are not clearly addressed.

  • 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?

    (1) Lack of critical experimental details, affecting reproducibility and credibility; (2) Insufficient statistical validation; (3) Incorrect references; (4) Two-center external tests were conducted, with significant improvement observed in external testing.

  • 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 addresses most of my main concerns. The authors have added SEN, SPE, 95% CIs, and DeLong tests, clarified the CT input and preprocessing pipeline, specified key variable dimensions, corrected the RECIST-related reference, and provided additional implementation details. These revisions substantially improve the statistical reporting, reproducibility, and clinical clarity of the paper. I still recommend that the final manuscript explicitly include these clarifications and acknowledge the remaining limitation of single-split evaluation.



Author Feedback

We thank all reviewers’ valuable comments. Here are our responses: Q1 for Meta&R1&3: Lacks SEN, SPE, 95%CI, and Delong test Tables1–3 now report SEN, SPE, 95%CI (1000 bootstrap iteration) for all models. Updated metrics for our model (AUC/ACC/SEN/SPE, with 95%CIs) are as follows. Internal testing: 0.848(0.788–0.904)/0.752(0.680–0.817)/0.742(0.638–0.846)/0.759(0.671–0.849), significantly outperforming TRL and HARM3 (both p<0.01, DeLong test). External testing: 0.828(0.787–0.868)/0.754(0.708–0.797)/0.780(0.719–0.843)/0.735(0.671–0.797), significantly outperforming LC-NICER, M2Fusion, TRL, and HARM3 (all p<0.01, DeLong test). Q2 for R1: Lacks the discussion of limitations Limitations on multimodal integration, lack of prospective validation, and generalizability to other tumor types have been added to the final version. Q3 for Meta&R1: Mechanisms of the TIVDM is unclear Disentanglement in the parallel 3D encoders is driven by asymmetric supervision. Both branches are concatenated and trained via BCE loss to extract response-relevant features. Within this subspace, the upper branch is constrained by a cross-time MSE consistency loss and a batch distance loss to learn time-invariant, response-relevant features while preventing collapse. A subspace orthogonality loss forces the lower branch to capture the response-relevant time-variant signals. Q4 for Meta&R2: Loss function weighting is unclear The weighting coefficients λ1/λ2/λ3/λ4 were determined via grid search to maximize the validation set AUC, yielding 0.4/0.2/0.2/0.2.Q5 for Meta&R2: Limits details of 3D encoders and TGM The 3D encoders are implemented as 3D Swin Transformers, adapted from the 2D Swin by upgrading all spatial operators to their 3D counterparts. The architecture uses an initial embedding dimension of 16 and a depth configuration of [2,2,2,2] (8 Swin Transformer blocks in total) for input volumes of 128×128×128.The TGM is a two-layer MLP with hidden layer W1∈ℝ^(128×1)(128 units) and output layer W2∈ℝ^(256×128) (256 units), matched to the feature channel dimension. Q6 for Meta&R2: Clarity on time between CT scans The median pre- and post-treatment CT intervals at the two centers were 8.4 weeks (range 4.6–29.6) and 8.0 weeks (range 4.6–24.0), aligning with the 6–9 week window typically used for IO response assessment, with deviations reflecting real-world scheduling variability. Q7 for R2: Potential for overfitting and data augmentation Multiple strategies to mitigate overfitting was applied. (i) Data augmentation via TorchIO(random flipping, random Gaussian noise); (ii) Regularization: encoder dropout, weight decay through AdamW, and LR decay during training. Q8 for Meta&R2: Patient histology Our dataset comprises exclusively patients with non-small cell lung cancer, as now clarified in the final version. Q9 for R2: Typo We have made the corrections. Q10 for R3: Figure 1 lacks aesthetic appeal and clarity Figure 1 has been redesigned in the revised manuscript to improve clarity and visual quality, including clearer module labels, more structured model feature flow, simplified data-flow arrows. Q11 for R3: The formulas and variables fail to specify variable dimensions The revised manuscript explicitly specifies dimensions and domains of all key variables and feature representations. X_pre, X_post∈ℝ^(D×H×W); Δt∈ℝ⁺; F_inv∈ℝ^(d); F̃_inv, F̃_var∈ℝ^(B×2d); TGM weights W1∈ℝ^(d’×1), W2∈ℝ^(2d×d’) with biases b1∈ℝ^(d’), b2∈ℝ^(2d) (d, d’=feature dimension; B=batch size). Q12 for R3: Reference errors exist This has been corrected in the revised version with the proper reference: Eur. J. Cancer 2009; 45(2): 228–247.Q13 for Meta&R3: CT Input specification Our model can automatically process whole-lung volumes (for single- and multi-lesion cases) without tumor segmentation, ROI annotation, or lesion detection. Preprocessing includes resampling to 1×1×1 mm³ spacing, lung-window normalization, automatic cropping to lung bounding box, and resizing to 128×128×128.




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.

    The paper presents a method to disentangle treatment-related dynamic changes from patient-specific static features in longitudinal CTs. The reviewers highlighted the impressive scale of the multi-center cohort and the promising performance gains observed in external validation as major strengths. However, several concerns must be addressed for a more informed decision on acceptance/rejection –> several methodological and reporting gaps are identified by R1 and R2.The authors must provide confidence intervals and additional metrics, such as sensitivity and specificity, to move beyond point estimates and confirm the statistical reliability of their results. The rebuttal should also clarify the technical mechanisms of the feature disentanglement module, specifically explaining how the parallel 3D encoders effectively distinguish between temporal variant and invariant features. Also, the reviewers would like the authors to provide transparency regarding the selection of loss function weights and the implementation details of the time-gating mechanism to ensure reproducibility. Finally, clinical ambiguities concerning the patient histology, the variance in scan intervals, and the specific nature of the input ROI must be resolved. Addressing these points will be essential.

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

    Reject

  • Please justify your recommendation.

    Though the authors have attempted to address the concerns in the rebuttal, I still have concerns that reman unresolved. In several instances within the rebuttal, the authors state that the reviewers’ concerns have been addressed in the final version of the manuscript. However, they have not provided the necessary explanations or revisions within the rebuttal text itself. Given that MICCAI policy strictly prohibits manuscript modifications during this phase, these claims cannot be verified. Comments on statistical validation/overfitting are not convincingly addressed



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.

    Authors have adequately answered questions from reviewers.



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 authors have adequately addressed all of the reviewers’ concerns. Accordingly, I recommend acceptance of the paper.



back to top