List of Papers Browse by Subject Areas Author List
Abstract
Non-small cell lung cancer (NSCLC) is one of the leading causes of cancer-related mortality, with lymph node metastasis serving as a critical factor in both prognosis and treatment decisions. Lymph node station (LNS) dissection is an essential procedure in the management of NSCLC patients; however, over-dissection may expose patients to unnecessary risks, while under-dissection could lead to undetected metastases. Despite its importance, predicting the exact metastasis status during surgery remains challenging. To address this challenge and meet the urgent need in clinical practice, this study presents the Deep Knowledge-infused Transformer (DKiT) model, designed to predict LNS metastasis in previously unexamined regions by capturing the relationships between LNSs. Furthermore, DKiT is augmented with clinical prior knowledge through a multi-stage infusion mechanism during the decoding phase, enhancing both model performance and interpretability. Additionally, we developed an AI-powered intraoperative decision support system based on DKiT, which provides real-time surgical recommendations informed by frozen pathology results. Experimental results show that DKiT achieves an AUC score of 0.812 for LNS-level metastasis prediction, outperforming other comparative methods. The clinical system achieves a recall of 0.930 and precision of 0.865 in the retrospective cohort collected from collaborating hospitals, highlighting its potential in guiding NSCLC treatment decisions. The source code is available at https://github.com/czifan/DKiT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3436_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/czifan/DKiT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaJie_Deep_MICCAI2025,
author = { Zhao, Jie and Chen, Zifan and Yang, Guangzhengao and He, Yijiang and Zhang, Li and Dong, Bin},
title = { { Deep Knowledge-Infused Transformer for NSCLC Lymph Node Station Metastasis Prediction: Development of an AI-Powered Intraoperative Decision System } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {168 -- 177}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents DKiT, a Deep Knowledge-infused Transformer designed to predict metastasis in unexamined lymph node stations (LNSs) of non-small cell lung cancer (NSCLC) patients during surgery. The model uniquely combines patient-specific partially known LNS status with multiple clinical prior knowledge graphs—capturing anatomic adjacency, rule-based transfer, and data-discovered metastasis pathways—via a multi-stage cross-attention mechanism within the Transformer decoder. The method is evaluated on a dataset of 919 patients and achieves state-of-the-art performance in LNS-level metastasis prediction (AUCLNS = 0.812) and surgical decision recall (up to 0.930), significantly outperforming baseline statistical, machine learning, and graph-based models. In addition, the authors propose an AI-powered intraoperative decision system that uses the model’s outputs to iteratively recommend the next LNS to dissect, offering real-time support for surgical planning.
- 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 paper reformulates LNS metastasis prediction as a partial-information node classification task, which realistically matches intraoperative settings where only a few LNS statuses are known. (2) It proposes a novel multi-stage knowledge infusion mechanism that dynamically integrates multiple clinical knowledge graphs into a Transformer decoder, improving prediction with limited input. (3) The model is tested on a large, real-world dataset (919 patients) and shows clear gains over statistical, machine learning, and graph-based baselines, especially in low-information scenarios. (4) The authors also implement a practical intraoperative decision system, making the method clinically actionable.
- 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 core architectural components—Transformer encoder-decoder and graph-based attention—are adapted from existing designs, and the novelty mainly lies in their combination and application rather than methodological innovation. (2) The knowledge graphs used for infusion (anatomic adjacency, rule-based, and data-driven) are not deeply analyzed or validated individually, and their construction lacks detailed justification or reproducibility. (3) The baseline comparisons are limited to relatively simple methods; stronger baselines such as Graph Transformers or memory-based models could better support claims of superiority. (4) The evaluation is entirely retrospective and single-center; no external validation or prospective study is conducted, which limits claims about clinical generalizability.
- 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.
- 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/2025/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 — could be 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 proposes a well-structured method for intraoperative LNS metastasis prediction using a knowledge-infused Transformer, with a clinically meaningful formulation and good performance on real data. The multi-stage knowledge injection is thoughtfully designed. However, the method relies on known architectures, baseline comparisons are limited, and—critically—the anonymous link provided contains no usable code or data, limiting reproducibility. With these issues addressed, the work would merit acceptance.
- 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 knowledge-infused transformer-based model aimed at improving the prediction of lymph node metastasis. In addition, the authors present an intraoperative decision support system built upon this model, evaluating it through experiments that simulate real-world surgical settings.
- 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 work presents a welcome alternative to the current rule-based intraoperative methods, which, despite being the standard of care, are relatively rudimentary.
- The concept of incorporating clinical knowledge through explicit representations at multiple stages of the deep learning pipeline is both novel and compelling.
- 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.
- Since the model is positioned as a real-time assistant in the operating room (OR), a direct comparison with experts in terms of both time and accuracy would strengthen the experimental results. At a minimum, the paper should report the approximate time typically taken by experts to complete the task and the average error rate observed in clinical practice.
- The paper would benefit from reporting the number of parameters for both the proposed model and the baseline models. This information is important, as lighter, more interpretable models that perform comparably well might be preferable in intraoperative settings.
- Although the paper mentions that the train, validation, and test splits were patient-based, it remains unclear how the model input was processed during training. Specifically, given that the model adopts a mask-and-predict strategy and backpropagates losses for known lymph node stations (LNSs), clarification is needed on whether backpropagation occurred once per patient. A description of how patients were batched during training would improve reproducibility.
- Similarly, more details are required regarding the input processing for the baseline models. For instance, for the graph-based model, how were the graphs constructed?
- While K1, K2, and K3 could be interpreted as intermediate representations from the graph encoder for knowledge infusion, their explanations introduce new concepts that had not been previously described. This section should be clarified to improve the interpretability of the ablation study results.
- The paper states that correlation analysis was performed using the initial location embeddings (L). If this is the case, the conclusions drawn from this analysis could be misleading. To better validate the claim that the proposed Transformer model captures complex LNS relationships, it would be more appropriate to conduct correlation analysis based on the final learned representations rather than the initial embeddings.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/en/REVIEWER-GUIDELINES.html
- Please ensure that all mathematical terms are defined when introduced to improve readability of the paper.
- 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 — could be 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 clinical problem and proposes a novel knowledge-infused transformer model for intraoperative decision support. However, several aspects require clarification, particularly regarding training procedures, baseline construction, and the interpretation of correlation analyses. Provided that the authors can satisfactorily address these concerns during the rebuttal, I would recommend acceptance.
- 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
The article proposes a decision-support system using a transformer-style architecture for predicting lymph node metastases specifically using a graph structure that updates itself when pathology information about different lymph nodes is included (via a separate transformer architecture). The goal is to have a largely flexible framework that can take advantage of clinically available information regarding partial solutions to the prediction problem (i.e. which lymph nodes are healthy versus those with metastases) rather than prediction purely from nothing.
- 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 clinical motivation behind the architecture is quite strong and one can readily see why the particular techniques were used.
- Figure 3 A shows the improvement from adding more knowledge about surrounding lymph nodes for the prediction of the remainder, which was the clinical goal of the framework.
- Statistical testing (via 95% confidence intervals in Table 1) shows the robustness of the methods with respect to the state-of-the-art.
- The authors should that analysis of the embeddings gives significant interpretational strength to the model, making it easier to trust.
- 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.
- Some more work could have been done in terms of integrating uncertainties and hypothetical situations as it is not only important to biopsy knowably positive lymph nodes but to also use biopsy to clarify lymph nodes that have an unknown status and could provide significant information about their surroundings.
- Although the networks are explained simply, there is significant implementation details that are lacking with no clear indication that the authors intend on making their code openly available.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
I would strongly suggest making the architecture open-source, preferable with parameter weights. If the underlying anonymised data could also be made open-source, that would significantly benefit the community.
- 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?
Overall, I think this is a good paper that presents a new machine learning method but does so in a way where the clinical problem drives the technical decisions. This gives the paper significant interest to people, especially in the CAI community, interested in new approaches to interaction machine learning architectures.
- 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 appreciate the reviewer’s valuable comments and will incorporate the suggested changes in the final version.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
The reviewers were uniformly positive on this paper, with however well-defined reservations. I urge the authors to take advantage of the thoughtful reviews to further improve and tune their work to better connect with the MICCAI audience.