List of Papers Browse by Subject Areas Author List
Abstract
Anthracotic carbon pigment deposition reflects cumulative inhaled particulate exposure and is only visible in histopathology, which is however invasive, limiting scalability for screening and longitudinal monitoring. We developed a CT-based radi-omics framework to non-invasively estimate pollutant burden using pathology-derived lung pollutant measure as reference. We extracted standardized radiomic features from six anatomical regions (tumor, whole lung, normal lung, and 5/10/15-mm peritumoral rings) and trained region-specific classifiers with nested cross-validation. Region-specific classifiers trained via cross-validated feature selection were integrated through weighted late-fusion. In an in-houe cohort (n=145), multi-regional fusion achieved AUC=0.733 (95% CI: 0.644–0.812), outperforming single-region models (best AUC=0.664). Normal lung parenchyma contributed most strongly (weight=0.40), with texture heterogeneity features dominating. External evaluation on a subset of National Lung Screening Trial (NLST) (n=194) demon-strated maintained discriminative performance under domain shift (AUC=0.609, 95% CI: 0.513–0.702). Subgroup analyses suggested the model is not driven by smoking status, confirming pollution-specific signal capture. This framework pro-vides a noninvasive, mechanism-anchored approach to quantify pollutant-associated phenotypes from routine chest CT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1327_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{LiYut_Multiregional_MICCAI2026,
author = { Li, Yutong AND Pan, Xiaoxi AND Balachandran, Abishek AND Behrens, Carmen AND Salvatierra, Maria Esther AND Solis Soto, Luisa Maren AND Young, Chingyi AND Yuan, Yinyin AND Wu, Chengyue},
title = { { Multi-regional CT Radiomics Fusion to Quantify Pathology-Derived Pollutant Burden in 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 work provides an alternative scheme to quantify lung pollutant burden in lung cancer, using routine chest CT scans instead of the standard pathology (invasive biopsies).
- 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 proposed framework was trained with an in-house cohort and tested with an external trial dataset, which tries to demonstrate its generalizability. -Based on the fact that classification models from six different regions were integrated, it allows to capture different patterns which correlate with biological understanding, for instance, diffuse parenchymal patterns and localized tumor-host interactions.
- 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.
-Based on the fact that region segmentations used to extract features were done manually, which could introduce inter-observer variability and be labor-intensive. It does not affect the performance presented in the manuscript, it could represent a limitation to be implemented in a real-world clinical workflow. -A good point for model training is that the cohort for that purpose was balanced in terms of the classes, however it is not the same for the external cohort, and it can be noted with the shift in precision-recall metrics. -This comment could be taken as contradictory with one of the strengths, but although I recognize the effort of the authors to use two different cohorts (different institutions, public/private, acquisition parameters), there is a notable drop performance between training and testing (from AUC of 0.733 in training to AUC of 0.609 in the external NLST dataset).
- 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
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 main reason for this decision is the notable drop performance between training and testing cohorts.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The main contribution is to develop a radiomics based model to assess pollution burden based on lung CTs.
- 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.Interesting question: the problem of determining a CT based lung pollutant index seems very interesting and potentially impactful to me. 2.Good methodological design and rigor: the paper is methodologically sound and the experiments are well designed (much more rigorous than typical conference papers). The design choices are well justified and also stress tested (for example, the sensitivity of results to choice of thresholds). The external validation is also performed quite rigorously – the same numeric LPI cutoff is used as the threshold for the binary labeling. Pretty comprehensive ablation studies and subgroup analyses have also been performed. I commend the authors for a well designed study. 3.The organization and writing are clear.
- 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.
Weaknesses: 1.Some important details are missing: For example, in the NLST data, each study has several series with different reconstruction kernels. Since the study uses radiomics features which can be quite sensitive to changes in reconstruction kernels, the authors need to mention/specify how/which series were chosen for analysis. This worries me a bit, since radiomics features often tend to not generalize well due to these factors, and the current AUC values are quite low as it is. 2.Manual segmentations seem to have been used in the analysis. However, that is not really scalable. I wonder how sensitive the findings are to variability of segmentations.
3.Instead of the dichotomizing the LPI to a binary classification problem, it might have been better to treat it as regression problem with continuous valued outputs. This is one point I disagree with the authors from a methodological standpoint.
- 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
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 study is quite well done and the methods are rigourous. The problem is also quite interesting in my opinion.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This study addresses a novel clinical problem by proposing a non-invasive assessment of pulmonary pollutant burden in lung cancer patients using a CT-based radiomics model. Individualized evaluation of pollutant exposure may provide an essential reference for lung health management and deliver significant clinical value for precise diagnosis and targeted treatment.
- 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) Building on previous research, this study defined a Lung Pollutant Index (LPI) based on information derived from pathological slice images. The proposed index may serve as a novel risk factor for the health management of pulmonary pollution status in general populations, especially among non-smokers. (2) Constructing a non-invasive CT-based model to estimate the LPI represents a feasible and population-friendly solution with great public health significance. (3) The validation and evaluation of this study were conducted on both internal datasets and publicly available datasets.
- 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 LPI metric has not been formally published and is currently under peer review. Therefore, the validity of this indicator requires further rigorous evaluation. (2) The stratification criteria for the high‑LPI and low‑LPI groups lack sufficient justification. It remains unclear whether the current grouping threshold possesses genuine clinical significance or merely represents an empirical cutoff optimized for favorable model performance; solid supporting evidence is absent. (3) The reported AUC values are still relatively moderate and suboptimal. (4) A unified LPI threshold was adopted to dichotomize both the internal private dataset and the public dataset into high‑ and low‑LPI subgroups. Nevertheless, the public dataset exhibits a substantially higher proportion of high‑LPI samples. This raises concerns regarding the rationality of such a unified cutoff: heterogeneity in pathological slide information across different patient cohorts may compromise the computational stability and generalizability of the LPI calculation.
- 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
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?
This study addresses a brand-new and promising clinical pain point with the potential to improve outcomes for lung cancer populations, which is commendable and impressive. The overall research design is relatively rigorous and comprehensive. However, the performance of the current model still needs improvement, and the stability and rationality of the predicted target (LPI) require further clarification and optimization.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and meta-reviewers for their constructive comments. We address the major concerns below.
1.Manual segmentation and scalability. We agree that manual segmentation may introduce inter-observer variability and limit scalability. In this proof-of-concept study, manual delineation was used to ensure anatomically accurate region definitions for pathology-anchored model development. However, the proposed CT-LPI framework is not intrinsically dependent on manual segmentation and can be coupled with automated or semi-automated lung, tumor, and peritumoral segmentation methods. We also note that the main finding is based on consistent multi-region trends rather than a single boundary-sensitive ROI, which may reduce dependence on any one manual contour.
2.External performance and domain shift. We agree that the external NLST AUC is moderate and lower than the internal AUC. We interpret this result as proof-of-concept rather than evidence of a deployment-ready clinical model. This is a challenging task because CT-LPI aims to predict subtle tissue-level pollutant burden from routine CT, rather than a visually obvious disease phenotype. The NLST cohort also differs from the internal cohort in case mix, disease stage, acquisition protocol, scanner settings, and reconstruction kernel. Despite this stringent external setting, the model remained above chance in NLST, supporting the presence of partially transferable pollutant-associated CT signal.
3.NLST series selection and reconstruction kernels. We thank the reviewers for emphasizing reconstruction-kernel heterogeneity. For NLST, we selected the CT reconstruction series referenced by the tumor segmentation and excluded cases in which tumor and lung masks referenced different series. The selected NLST series had heterogeneous kernels, whereas the internal cohort used LUNG kernels. Thus, sharp-equivalent lung kernels represented only a subset of NLST. We agree that this likely contributed to the external performance drop. We will clarify the series-selection procedure and reconstruction-kernel limitation.
4.LPI definition, dichotomization, and threshold transfer. LPI is not an arbitrary label; it is derived from quantification of anthracotic carbon pigment deposition in lung tissue and therefore provides a tissue-level measure of pollutant-associated burden. Our goal was not to establish LPI as a finalized clinical biomarker, but to test whether CT contains imaging features predictive of tissue-resident pollutant burden. We formulated the task as classification because LPI values are small and occupy a narrow range, where precise continuous prediction may be unstable in the current cohort size. The high/low cutoff was pre-specified as the 60th percentile in the internal cohort, corresponding to LPI ≈ 0.001, to provide sufficient representation of the high-LPI subgroup while avoiding extreme tail grouping; it was not optimized for model performance. The same numeric cutoff was applied to NLST to avoid re-optimizing the threshold on the external set. We agree that the higher high-LPI proportion in NLST may reflect both true cohort differences and pathology-side heterogeneity, including slide availability, tissue sampling, staining, and scanning.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents a non-invasive approach to quantify pulmonary pollutant burden via a multi-regional CT radiomics framework, addressing a significant unmet clinical need - Though technically not novel, the work has implications in lung health management. Reviewers commended the study’s methodological rigor, specifically the well-justified feature fusion to capture complex biological patterns associated with pollutant exposure. The evaluation is strengthened by the inclusion of both an internal cohort and an independent external validation using the NLST dataset. Overall, the work provides a population-friendly alternative to invasive biopsies. The camera ready can focus on addressing the minor weaknesses raised
