List of Papers Browse by Subject Areas Author List
Abstract
In clinical routine, expert radiologists localize peripheral or central features in lung tissue, including vessels, bronchi, and other structures, to identify and categorize pulmonary nodules (PN). However, current computational PN detection methods rely on standard convolutional and transformer-based models without exploiting this irregular anatomical and contextual coherence within the parenchyma, which may be key to detecting nodular lesions in early stages. Emulating expert analysis, this work proposes a hierarchical, geometric graph representation that considers non-local and non-rigid relationships among lung regions. Using this representation, we learn a differentiable soft cluster assignment via a graph pooling operation, converging into a binary background/nodule representation. We encode each node’s features with a second-order embedding to capture local inter-feature covariances, enabling spatially and contextually coherent detection. Moreover, we introduce an orthogonality regularization to enforce dissimilarity between the two embeddings in the grouping-matrix space. We evaluated our method on the LIDC-IDRI dataset, achieving a CPM of 0.51, which outperforms baselines such as YOLOv5, Faster R-CNN, DETR, and Grounding-DINO. In a controlled subset designed to validate the contribution of lung structures to the localization task, our method achieved a notable CPM of 0.74 (IoU=0.1), supporting the initial hypothesis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6009_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://gitlab.com/luis.guayacan1/miccai_2026-nodulesdetection
Link to the Dataset(s)
LIDC-IDRI dataset: https://www.cancerimagingarchive.net/collection/lidc-idri/
BibTex
@InProceedings{GuaLui_SpatiallyAware_MICCAI2026,
author = { Guayacán, Luis AND Leal, Santiago AND Patiño, Diego AND Martinez, Fabio},
title = { { Spatially-Aware Graph Geometric Representation for Lung Nodules Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a graph-based framework for pulmonary nodule detection that models lung parenchyma as a graph of anatomically meaningful superpixels, aiming to better capture the irregular and contextual nature of lung tissue compared to standard rigid grids. To capture local textures, nodes are encoded with second-order covariance descriptors and processed via a hierarchical differentiable pooling (DiffPool) network to converge into a simplified binary graph representing “nodule” and “background” clusters. Evaluated on the LIDC-IDRI dataset using 2D 3-slice stacking, the method achieves a CPM of 0.505, outperforming several standard 2D object detectors such as YOLOv5, Faster R-CNN, and Grounding-DINO.
- 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.
Moving away from rigid grid-based CNNs or fixed-patch transformers toward a superpixel graph is an anatomically grounded and highly logical shift. It natively addresses the morphological heterogeneity of the lung, where nodules frequently mimic or touch irregular structures like vessels and the pleura.
The use of second-order covariance matrices to encode inter-channel dependencies effectively enriches the local node embeddings before message passing. Furthermore, introducing an orthogonality regularization term is a mathematically sound way to prevent representation collapse during the DiffPool clustering phase, directly addressing the severe class imbalance between nodule and non-nodule tissue. The ablation study in Table 2 convincingly isolates the individual contributions of the covariance descriptors and the orthogonality regularization, proving that both are necessary for optimal performance.
- 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 framework relies on a 2.5D approach (stacking 3 axial slices), and the authors evaluate it strictly against 2D object detectors. The authors themselves acknowledge that current benchmarks heavily favor 3D CNNs and 3D attention mechanisms to leverage full volumetric continuity. Comparing a highly specialized medical algorithm exclusively against generalized 2D vision models (like YOLOv5 or DETR) limits the ability to assess true SOTA clinical performance. Can the authors discuss how this 2D graph approach compares to standard native 3D detection architectures?
While DiffPool provides excellent hierarchical coarsening, compressing the entirety of the complex lung graph into strictly 2 final nodes (background and nodule) may force a suboptimal “hard” bottleneck. The lung background is highly heterogeneous (containing air, blood vessels, bronchi, and pleura). Forcing all non-nodular superpixels into a single unified background cluster might destroy granular semantic context before the final attention mechanism. Could the authors clarify why a binary bottleneck was chosen over a multi-class semantic clustering approach?
The authors note in their conclusion that the reliance on SLIC superpixels and second-order feature computations is “computationally expensive”. Given that standard single-stage anchor-based detectors are heavily optimized for speed, can the authors provide inference time metrics (e. g. , seconds per scan) to contextualize the trade-off between this graph representation and baseline models?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a creative, scientifically grounded approach to lung nodule detection by leveraging superpixel graphs, second-order features, and differentiable pooling. The primary strength lies in its conceptual shift away from rigid grids to capture irregular parenchyma topology, which is a strong representation learning contribution. However, the reliance on 2D evaluation baselines in a field dominated by 3D volumetric models, combined with unanswered questions regarding computational overhead, weakens the empirical claims.
- 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 utilizes the ability of Graph Neural Network in understanding non-grid structures to incorporate contextual information of lung region for better pulmonary nodule detection. Using superpixel-based graph on the input segmented lung region, node embedded graph is formed. To overcome imbalanced nodule and non-nodule nodes of the graph, DiffPooling with orthogonal regularization is used to reduce the graph into binary node graph. With the help of features from both initial and binary graphs, MLP is used for predicting bounding box and confidence scores of lung nodules. The work is evaluated in standard pulmonary nodule LIDC-IDRI dataset.
- 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 proposed work has applied GNN in a novel way over the superpixels of lung region to incorporate contextual information. 2.This technique can be applied to other nodule/lesion of other organs or modalities for detection as well. 3.The proposed work is evaluated in standard pulmonary nodule LIDC-IDRI dataset.
- 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 proposed work is only compared with the state-of-the-arts in 2D object detection but not with the existing state-of-the-arts in lung nodule detection such as [1], [2]. 2.The work does not explicitly mention how the two node graph \mathcal{G}^{(f)} = (A^{(f)}, \boldsymbol{X}^{(f)})from the reduced graph \mathcal{G}^{(l+1)} = (A^{(l+1)}, \boldsymbol{X}^{(l+1)}) using DiffPool. 3.The use of multiple networks(Simple Linear Iterative Clustering, GNN, DiffPool, MLP) made the end-to-end process computationally expensive. 4.There are not enough results to validate the authors claim that “contribution of contextual lung regions to the lesion localization task”. That is, the high performance of the proposed work over other compared works in pulmonary nodule detection is not only because of incorporation of contextual information of lungs.
[1] Tang, Hao, Chupeng Zhang, and Xiaohui Xie. “Nodulenet: Decoupled false positive reduction for pulmonary nodule detection and segmentation.” International conference on medical image computing and computer-assisted intervention. Cham: Springer International Publishing, 2019.[2] Ding, J., Li, A., Hu, Z., Wang, L.: Accurate pulmonary nodule detection in computed tomography images using deep convolutional neural networks. In: Descoteaux, M., Maier-Hein, L., Franz, A., Jannin, P., Collins, D.L., Duchesne, S. (eds.) MICCAI 2017.LNCS, vol. 10435, pp. 559–567.Springer, Cham (2017).https://doi.org/10.1007/978-3-319-66179-7 64
- 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
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?
Despite the lack of comparison of the proposed method with the state-of-the arts in lung nodule detection, the method is novel in usage of GNN over superpixels. Making the model lightweight would add some more weightage to the framework. Further more evaluations are needed to support the claim “contribution of contextual lung regions to the lesion localization task”
- 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.
Reviewer is satisfied with the author’s rebuttal. In specific, the justification of computational complexity comparison of purposed work with Yolov5s, Faster-RCNN and RTDETR. Despite receiving critical comments, the work used novel idea which can impact other CAD detection using different modalities and organs.
Review #3
- Please describe the contribution of the paper
1.The paper proposes a graph-based method for lung nodule detection. The method models the lung as a superpixel graph with anatomically meaningful spatial and contextual relationships, then uses hierarchical differentiable pooling plus second-order feature encoding and an orthogonality regularizer to separate nodule from background regions. The experiments on LIDC-IDRI show improved pulmonary nodule detection performance over several baselines.
- 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 method represents the lung as a superpixel graph and explicitly models non-local anatomical and contextual relationships between lung regions instead of treating CT slices as regular grids or patches. The novelty is to learn a graph that progressively form semantically meaningful groups while encouraging clean separation between background and nodule representations.
- 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 comparison is not up to date. The paper mainly benchmarks against older or generic 2D detectors, without comparing to stronger recent pulmonary nodule detection methods, especially medical-specific approaches. 2.The paper attributes the gains to anatomical context and graph reasoning, but the ablation study does not clearly separate the effects of the graph representation, superpixel construction, pooling strategy, and detection head.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well motivated, proposes a reasonably novel task-specific combination of graph-based anatomical modeling, second-order features, and hierarchical pooling, and shows clear gains over several detector baselines on LIDC-IDRI. However, the experimental setting is narrower than the clinical problem the paper aims to address, the comparison set is not sufficiently up to date, and the current evidence does not fully prove the benefit of the proposed graph-based contextual modeling. The authors should clarify their baseline choices, better position the work against recent nodule detection methods, and explain more clearly why the current experiments are sufficient to support the claimed value of the proposed representation.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal clarifies design choices and computational cost, but the main concern remains unresolved. The paper only compares against generic 2D detectors (YOLOv5, Faster R-CNN, RT-DETR, Grounding-DINO) and not against strong, medical-specific pulmonary nodule detection methods or 3D/volumetric models. As such, the reported improvements cannot convincingly establish the effectiveness of the proposed graph-based representation. Ablation studies also do not clearly isolate the contribution of the graph context versus other components. While the idea is interesting, the empirical evidence is insufficient for acceptance. The comparison remains largely limited to generic 2D object detectors, while strong pulmonary nodule detection methods, particularly medical-specific and 3D approaches, are not adequately included. Although the authors explain that their goal is to isolate the benefit of graph reasoning under matched 2D/2.5D input conditions, this makes the performance claims narrower than the paper’s broader positioning as a lung nodule detection method. The preliminary 3D result mentioned in the rebuttal is encouraging but cannot substitute for a complete, well-described comparison in the submitted paper. Moreover, the ablation study still does not fully isolate the effect of the proposed graph-based contextual modeling from other components such as superpixel construction, covariance descriptors, DiffPool, and the prediction head. As a result, the central empirical claim remains insufficiently supported. Therefore, despite the novelty of the idea, I maintain a reject recommendation.
Author Feedback
We thank the reviewers for their valuable comments and feedback.
Comparison with 3D Nodule Detectors. (R1,R2,R3). We agree that 3D pulmonary nodule detectors represent the current state of the art. However, their performance is largely supported by volumetric/inter-slice continuity, often extracted from cropped 3D patches. In our work, the goal is to assess whether graph reasoning can exploit whole-image anatomical context, where the model must distinguish nodules from vessels, pleura, and other lung structures through non-local relationships across the full slice. Therefore, we evaluated our method in a 2D/2.5D regime and compared it with detectors operating under matched input conditions, avoiding attribution of gains to native 3D continuity. We will clarify this strategy in the revised manuscript. Moreover, to address the reviewers’ concerns, we performed a preliminary 3D extension, constructing graphs from a 3D autoencoder, following a similar graph-based framework, achieving a competitive CPM of 0.83.This preview of our future work suggests that the proposed representation can also benefit from inter-slice continuity while preserving spatial morphological context.
Two-node DiffPool Graph. (R1, R2). The two-node bottleneck is designed to provide a task-driven global representation, not to replace local features. Given \mathcal{G}^{(l)}=(A^{(l)},X^{(l)}), the model computes node embeddings Z^{(l)} and soft assignments S^{(l)}, producing X^{(l+1)}=S^{(l)T}Z^{(l)} and A^{(l+1)}=S^{(l)T}A^{(l)}S^{(l)}. In the last pooling layer, S^{(L-1)}\in\mathbb{R}^{N_{L-1}\times2}, yielding \mathcal{G}^{(f)}=(A^{(f)},X^{(f)}) with two task-driven latent nodes. Although the lung is heterogeneous, the initial superpixel-level embeddings are preserved and combined with this global representation. Thus, the final graph serves as a discriminative regularizer, enriching the node embeddings fed to the detection head for nodule localization. A multi-cluster formulation could model finer subclasses, but in the current framework it would require additional labels. The reviewer’s suggestion motivates future self-supervised strategies to learn finer anatomical subclasses without dense semantic labels.
Computational Cost. (R1,R2). In our method, the computational cost is concentrated in the initial SLIC-based graph and covariance-descriptor computations, which constitute the main runtime bottleneck. After this stage, the model operates on a compact, sparse graph rather than on the full pixel grid. In our implementation, inference takes 0.034 s per 2.5D image. Under the same setting, YOLOv5s, Faster R-CNN, and RT-DETR take 0.0081 s, 0.0596 s, and 0.0376 s, respectively. We highlight that a) our method’s inference time is competitive with all the baselines, even surpassing Faster R-CNN and RT-DETR, and b) we overstated the computational complexity claim of the superpixels strategy.
Ablation and Contextual-graph Claim. (R2,R3). We agree that the current ablation does not fully disentangle all components of our proposed strategy. We will clarify in the camera-ready that the results support the graph-based contextual representation as a whole, rather than isolating the specific contribution of contextual lung regions. Still, we believe that our experiments in Table 2 show an important ablation of our method, i.e., removing covariance descriptors or orthogonality regularization degrades performance. These results indicate that the gains are not solely due to the final prediction head. Moreover, removing the graph representation essentially reduces the model to a YOLO-style grid-based detector, which is indirectly covered in the Table 1 baseline. We will consider a more exhaustive ablation in future work.
Reproducibility and Ethics. (R1,R2). We will release the code upon acceptance and include additional implementation details in the camera-ready. Ethical approval is not required as the dataset is publicly available.
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.
mixed reviews.
- 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.
There is 1 Accept and 1 Reject after rebuttal, while R1 did not update the final recommendation and remained positive. Reviewers generally found the graph-based formulation interesting and well motivated. However, the main concern remains the experimental validation. The paper is positioned as a lung nodule detection method, yet comparisons are largely limited to generic 2D detectors, making it difficult to assess the value of the proposed representation relative to stronger medical-specific and 3D approaches. While the rebuttal provided a preliminary 3D result, according to the MICCAI rebuttal policy, new experimental results cannot be considered as evidence for acceptance. More importantly, the need for such additional results reinforces the concern that the submitted experimental validation is not sufficient to fully support the paper’s claims. I therefore recommend rejection.
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.
Reviewers appreciated the novelty and partially found the rebuttal helpful, and the rebuttal addressed key concerns on computational cost and design choices. Despite the concern about limited comparisons, the overall strengths in methodological innovation and clarified contributions outweigh the remaining weaknesses, supporting acceptance.
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.
This paper proposes a graph-based framework for lung nodule detection. The reviewers generally recognized the novelty of the graph-based representation and the potential value of incorporating anatomical context. While the paper would benefit from stronger comparisons with medical-specific 3D nodule detection methods and more extensive ablation analyses, the proposed methodology is sufficiently novel, technically sound, and supported by encouraging experimental results. Overall, the strengths outweigh the remaining limitations, and I recommend Accept.