List of Papers Browse by Subject Areas Author List
Abstract
Imageless Total Knee Arthroplasty (ITKA) requires real- time acquisition of patient-specific anatomical landmarks without pre- operative CT or MRI, posing a significant challenge for intraoperative planning. Despite utilising 3D knee reconstruction algorithms, modern ITKA systems continue to struggle with accurate knee shape estimation and the anatomical axis alignment - both prerequisites for reliable ITKA planning. We present SMART-KNEE, a cascaded anatomy-aware recon- struction framework that generates patient-specific Femoral (F) and Tib- ial(T) bone morphology from a minimal set of intraoperatively collected anatomical landmarks, without preoperative imaging. In the first stage, a Topology-Aware Equivariant Graph Neural Network (EGNN) predicts the complete anatomical landmark set from four intraoperative anchor landmarks, enforcing topological consistency and anatomical plausibility under under E(3)-equivariance (invariance to 3D euclidean transforma- tions). In the second stage, an Anatomy-Aware GAN reconstructs com- plete patient-specific bone surfaces by fusing the EGNN-predicted land- marks with partial intraoperative surface point clouds. SMART-KNEE achieves precise landmark localization error (F: 2.28 mm ± 1.68 mm; T: 1.65 mm ± 0.83 mm). Test cohort surface reconstruction yields Dice F- 87%, T-92.477%, Hausdorff Distance F-5, T-4.3 mm and ground truth
- generated knee distances of 1.89 mm (F) and 2.87 mm (T), enabling
accurate patient-specific morphology retrieval for intraoperative ITKA
planning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6533_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/6533_supp.zip
Link to the Code Repository
https://github.com/IGRS-Imaging/SMART_KNEE
Link to the Dataset(s)
https://zenodo.org/records/20726242
BibTex
@InProceedings{RajDur_SMARTKNEE_MICCAI2026,
author = { Rajasekar, Durga AND S, Swetha Lakshmi AND M R, Vishnu AND Maik, Vivek AND Lakshmanan, Manojkumar AND Sivaprakasam, Mohanasankar},
title = { { SMART-KNEE: Sparse-Morphology Based Anatomical Reconstruction Using Topology-Aware Networks for Imageless Total Knee Arthroplasty } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The study presents SMART-KNEE, a cascaded DL framework for reconstructing 3D patient-specific femoral and tibial models during imageless total knee arthroplasty. There are two steps in the pipeline: first, an E(3)-equivariant graph neural network predicts a full set of anatomical landmarks from four intraoperative anchor points; second, an anatomy-aware GAN fuses these landmarks with a sparse intraoperative point cloud to reconstruct the dense bone surface. The authors validate the method on a physical phantom, achieving sub-3mm accuracy, and contribute a new dataset of 361 annotated bone models
- 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 technical formulation is sophisticated and well tailored to the surgical usecase. Using an E(3)-equivariant graph neural network is a strong design choice, as it ensures the network can accurately predict landmarks regardless of the arbitrary coordinate frame transformations inherent to intraoperative optical tracking. The system demonstrates impressive accuracy, significantly outperforming baselines like VN-EGNN in landmark prediction, and achieving surface reconstruction errors well below the 3 mm threshold generally accepted for clinical viability in TKA. Splitting the pipeline into a sparse structural prediction and dense morphological generatio elegantly constrains the generative model, ensuring anatomical plausibility.
- 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 study predicates its clinical narrative heavily on the assumption that robotic or highly navigated systems (which require 3D models) are the definitive standard of care. However, conventional TKA remains the standard in many regions, making the clinical motivation somewhat unstable. Framing this as a mandatory solution for the “standard” workflow is slightly overstated. While the methodology is great, the validation is restricted to in-silico metrics and a pristine physical phantom. In a real world intraoperative setting, acquiring the anchor points and sparse point clouds is complicated by soft tissue, cartilage thickness, blood, and surgeon variability. It remains unclear how robust the system is to noisy input data derivd from a real patient
- 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.
(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 study presents a great and novel methodology (combining EGNNs and GANs for a distinct clinical problem) and provides a potentially valuable dataset. The rating is currently a weak accept due to the gap between the mathematical/phantom validation and the realities of the clinical workflow, as well as the somewhat overstated clinical motivation. If the authors can adequately clarify the system’s robustness to realistic intraoperative noise and inference times in the rebuttal, this paper would be a good addition to the MICCAI program.
- 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.
My recommendation for acceptance is justified by the authors’ comprehensive and convincing rebuttal, which directly resolved the most critical concerns raised during the initial review. A key factor was the clarification that the validation was conducted on physical 3D printed phantoms using clinical-grade optical tracking rather than being purely in-silico, which effectively proves the system’s robustness against realistic intraoperative noise and operator variability. Additionally, the authors provided a strong technical argument regarding why traditional baselines like NeRF or diffusion models are unsuitable for the extreme sparsity of surgical point clouds, while simultaneously demonstrating that their approach significantly outperforms traditional statistical shape models by requiring fewer landmarks and avoiding mean-biased outputs
Review #2
- Please describe the contribution of the paper
This paper proposes SMART-KNEE, a cascaded framework for reconstructing patient-specific knee anatomy in imageless total knee arthroplasty (ITKA) from sparse intraoperative data. The method consists of: 1) A topology-aware E(3)-equivariant graph neural network (EGNN) that completes the full anatomical landmark set from only four intraoperative anchor points, enforcing geometric consistency and anatomical plausibility; 2) An anatomy-aware generative adversarial network (GAN) that reconstructs complete femoral and tibial bone surfaces by fusing predicted landmarks with sparse intraoperative surface point clouds. The framework eliminates the need for preoperative CT/MRI while achieving clinically acceptable accuracy, enabling real-time intraoperative planning. Additionally, the authors release a dataset of 361 annotated knee models.
- 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.Structured decomposition via anatomy-aware intermediate representation. The primary strength of this work lies in its formulation of imageless reconstruction as a structured two-stage problem, where anatomical landmark completion serves as an explicit intermediate representation between sparse intraoperative observations and dense surface reconstruction. This design is conceptually meaningful and non-trivial, as it introduces a clinically interpretable and geometrically constrained representation that reduces ambiguity in downstream generation. Unlike end-to-end implicit reconstruction approaches, this decomposition leverages anatomical priors in a modular manner, improving both stability and interpretability. 2.Topology-aware E(3)-equivariant modeling for landmark completion. A key methodological novelty is the integration of topology-aware graph modeling with E(3)-equivariant neural networks for anatomical landmark prediction. By explicitly encoding anatomical structure as a graph and enforcing equivariance under rigid transformations, the method ensures geometric consistency across varying poses and sparse inputs. This approach goes beyond conventional regression or statistical shape models by enabling structured propagation of anatomical constraints, which is particularly effective for recovering unobserved or geometrically distant landmarks. 3.Dual-conditioned generative reconstruction with anatomical constraints. The proposed reconstruction framework introduces a dual-conditioning mechanism that integrates local geometric evidence (partial point clouds) with global anatomical structure (predicted landmarks). This design provides a more controlled and anatomically grounded generative process compared to purely data-driven shape completion methods. Additionally, constraining the generator within a pretrained latent geometry manifold enhances realism and stability, representing a meaningful step toward reliable patient-specific reconstruction in clinical environments. 4.High clinical relevance with imageless and low-input requirements. The work demonstrates strong translational potential by achieving clinically acceptable reconstruction accuracy without reliance on preoperative CT or MRI. This significantly reduces the dependency on imaging resources while maintaining geometric fidelity, addressing a critical limitation in current ITKA systems. The ability to operate under minimal intraoperative input, combined with end-to-end validation, highlights the practical value of the approach for real-world surgical navigation scenarios.
- 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.Moderate level of methodological novelty with room for clearer differentiation. The methodological novelty of the framework is moderate and could be better articulated. While the integration of a topology-aware EGNN with an anatomy-aware GAN is well-motivated and tailored to the ITKA scenario, the core components themselves are based on established architectures. As such, the contribution is primarily reflected in the system-level design and problem formulation rather than in fundamentally new modeling techniques. To strengthen the novelty claim, the authors could more explicitly highlight what new capabilities emerge from this integration, for example through targeted ablation studies demonstrating the necessity of topology-aware modeling, or by providing theoretical or empirical analysis showing advantages over alternative architectural combinations. 2.Implicit reliance on learned statistical priors. Despite being framed as an imageless reconstruction approach, the method still relies on implicit statistical priors learned from data. The GAN-based reconstruction stage encodes population-level anatomical distributions within its latent space, effectively replacing explicit statistical shape models (SSMs) with learned priors. This raises a conceptual concern regarding whether the method truly overcomes the limitations of SSM-based approaches or simply shifts the dependency to a different representation. The authors could address this by providing experiments that evaluate generalization beyond the training distribution, such as testing on out-of-distribution anatomies or reduced training data settings, and by explicitly comparing the behavior of the model with classical SSM-based methods in terms of adaptability and bias. 3.Idealized assumptions on intraoperative input quality. The method relies on assumptions about the quality and distribution of intraoperative inputs that may be optimistic in practice. It assumes the availability of well-distributed anchor landmarks and relatively structured sparse surface point clouds, whereas real surgical data may be significantly noisier, incomplete, or unevenly distributed due to occlusion and acquisition constraints. This limits confidence in the robustness of the approach under realistic conditions. To improve this aspect, the authors could include robustness experiments by simulating noise, missing landmarks, or partial point clouds, and analyze performance degradation under such conditions to better reflect clinical applicability. 4.Lack of comparison with stronger or more recent baselines. The experimental evaluation lacks comparisons with several strong and relevant baselines from recent literature. In particular, methods based on diffusion models, implicit neural representations (e.g., NeRF, occupancy networks), or modern shape completion frameworks are not considered. Without such comparisons, it is difficult to assess the relative advantage of the proposed approach in the broader context of contemporary 3D reconstruction methods.
- 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.
(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?
1.Moderate level of methodological novelty with room for clearer differentiation. 2.Idealized assumptions on intraoperative input quality. 3.Lack of comparison with stronger or more recent baselines.
- 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.
After reading the rebuttal, I maintain a reject recommendation. The response does not sufficiently resolve my main concerns regarding clinical validity, novelty, and evaluation rigor. Although the authors clarify that the phantom evaluation was performed on 3D-printed CT-derived bone models with optical tracking, this remains a pre-clinical phantom study rather than evidence of robustness in real intraoperative, cadaveric, or patient-specific TKA settings. The claimed “clinical readiness” therefore remains overstated. Methodologically, the rebuttal argues for system-level novelty, but the framework is still mainly a combination of known components (EGNN-based landmark completion and GAN-based shape reconstruction) with limited evidence that each design choice is essential. The baseline comparison also remains insufficient: comparisons are made mainly against generic EGNN variants, while direct and fair comparisons with established statistical shape modeling, sparse point-cloud reconstruction, or recent implicit/generative reconstruction methods are not convincingly provided. Finally, several key concerns are deferred to future work, including systematic noise injection, missing-landmark analysis, partial-occlusion robustness, and broader validation. Thus, the rebuttal improves clarification but does not materially strengthen the submitted evidence to meet MICCAI acceptance standards.
Review #3
- Please describe the contribution of the paper
This paper presents a new method for a well-known problem in Total Knee Arthroplasty: the generation of patient-specific femoral and tibial bone surfaces from a minimal set of intraoperatively collected anatomical landmarks without preoperative imaging. It describes two deep learning networks that leverage on the specific anatomy and on topological consistency. The method is demonstrated on in-vitro experiments.
- 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 method may provide a time-effective and robust method for a well-established problem
- the emphasis on a minimal set of intraoperatively collected landmark points is important
- The Topology-Aware Equivariant Graph Neural Network and Anatomy Aware GAN combination is novel
- The experimental validation is sound and thorough
- 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 entrire reconstructed bone surface is not necessarily the target for the evaluation. Rather, the regions of interest for TKA surgery
- Limited test set
- Minor issues: writing style. The use of the letters F and T without stating that they refer to their full names once (Femur, Tibia). Starting sentences with a reference: “[8] reconstructed the complete F..” While I understand that this is for the purposes of manuscript length, it is ungrammatical and makes the paper harder to read.
- ” axis deviation below 3.5” – indicate the unit: millimeters, degrees?
- state what is E(3).
- 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 clinical relevance of the problem, novelty of the combined method and the experimental 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 have clarified most of the weak points in the rebuttal, which is sharp and well-written. I believe the paper can be accepted with the suggested revisions.
Author Feedback
(1)Meta-Review & R2.7.2 Shape-model-based reconstruction methods [1,12] share two structural limitations (i) Reconstruction is confined to a fixed statistical shape space (linear PCA subspace or deformable template), producing mean-biased outputs for typical morphologies (ii) Fitting requires 15–30 sparse landmark correspondences vs. SMART-KNEE’s 4 anchors. The L_patch coverage loss anchors reconstruction to observe intraoperative geometry, making the manifold a regulariser rather than a mean-shape generator absent in these methods. This is consequential as sparser inputs raise susceptibility to mean-biased reconstruction, propagating error into implant sizing and alignment. (iii) SMART-KNEE’s topology-aware conditioning preserves patient-specific fidelity under this sparse regime (mean T-ASD 1.69 mm vs. 3 mm threshold [18]). SSM infer anatomy by fitting a learned statistical shape space to sparse observations, reporting single-bone partial-surface error (Nolte et al. [12]: 2.60–2.95 mm, femur) or mechanical-axis deviation below 3.5°, whereas SMART-KNEE reconstructs complete F/T morphology from four anchors plus local patches (landmark RMSE 0.25/0.36 mm, ASD 2.4/1.7 mm F/T). Table 2 isolates the gain to prove that SMART-KNEE outperforms VN-EGNN [15] and EquiPPIS [14] by 4–6×. (2) Meta-Review, R1 & R2.7.3 Phantom validation and intraoperative noise.We correct a mischaracterisation , the evaluation (Fig. 5) is not in-silico. The phantom is a 3D-printed physical bone model from a real patient’s CT-segmented STL, complemented by 3D-printed models spanning ±2σ of the dataset’s morphological range (small/large condylar geometries, varus/valgus variants). Anchors and surface patches were acquired with a clinical-grade optical tracker and probe under realistic probe-tip noise, operator variability, and coordinate-frame perturbations not idealised input confirming robustness under practical acquisition conditions. Ground truth is CT-derived. Physical phantom evaluation with optical tracking is the established pre-clinical for surgical navigation [5,16,18] quantifying accuracy under clinical conditions prerequisite before cadaveric trials. Sub-millimetre accuracy on physically fabricated, optically tracked models establishes algorithmic validity necessary for translational progression. Anchor sensitivity analysis further quantifies robustness: errors reach 2.28/1.65 mm under suboptimal distributions vs. 0.25/0.36 mm under standard protocol. Systematic noise injection experiments (missing landmarks, partial occlusion) are a planned extension. (3) R2.7.1 & R2.7.4 Novelty and baselines. System-level novelty resides in three jointly necessary design choices Four sparse landmarks as a clinically interpretable anatomical intermediate, topology-aware E(3) equivariant chirality-conditioned completion and dual-conditioning using both global landmarks and intraoperative surface patches. Diffusion-based completion and implicit neural representations (NeRF, occupancy networks) assume 512–2,048 uniformly sampled points SMART-KNEE operates from ~70–100 points in a single surgical window 10–50× sparser. None target sparse landmark-conditioned reconstruction under E(3)-equivariance with anatomical graph constraints adaptation constitutes independent work. (4)R1 Imageless robotic/navigated TKA is a growing workflow [2,3,5,11,16] motivating SMART-KNEE, without implying conventional TKA is obsolete. (5) R3 Minor issues. F (Femur)/T (Tibia) defined at first text use, E(3) defined as equivariance to 3D rotation, translation and reflection “°” to “axis deviation 3.5°” rephrase reference-led sentences Nolte et al. [12] and add metrics over clinically critical resection zones. (6) R1 and R2 Reproducibility. Upon acceptance, we will release the annotated dataset (361 F/T models with landmark annotations), preprocessing pipeline, training code, and pretrained weights for full reproducibility.
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 submission received mixed reviews with one reviewer recommending acceptance, based on novelty of the method and significance of dataset, and two reviewers recommending rejection. The authors should address in the rebuttal the issues raised by the reviewers. In particular I would like to hear about the following: There are currently several navigation/robotic systems in the field for TKA that are imageless and leverage SSM, many times constrained by anatomical landmarks, to obtain a 3D model of femur/tibia. As referred by R2 (point 7.2), and similarly to SSM, the disclosed model also relies on learned statistical priors. Thus, what are the effective benefits with respect to SSM that has proven results in the field for many years, and why should the reader believe on these relative benefits given that the paper only provides results in phantom?
- 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.
The recommendations remain splitted with two reviewers suggesting acceptance (R1,R3) and one reviewer supporting rejection (R2). I am more aligned with this last recommendation because of the following: SSM is a well established approach that is the cornerstone of several navigation and robotics systems in the feld guiding the execution of thousands of TKA procedures every year (e.g. the Navio/CORI robot). In order to accept the work I would expect reasonable evidence that the proposed approach can outperform SSM in a clinically meaningful manner. The generation of such evidence does not necessarily require bioskills in “ex-vivo” specimens and can eventually be accomplished through experiments as the one of Fig 5, but with multiple phantoms that represent the variability of bone shapes, or through simulated acquisiton of landmarks and surface points in a vast annotated database of arthritic knee segmented from CTs or MRI. In my view the important is to compare the proposed approach againts the baseline SSM method in a representative number of bone F and T shapes using a clinical meaningful metric (ASSD, error in bone cut placement, etc). The authors only argue that conceptually SSM only produces “mean-biased outputs for typical morphologies” and provide a single experiment in phantom without any comparison with other methods that only shows that the their method can work in practice. This is clearly insufficient and for this reason my recommendation is to reject the paper
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.
I am aligning with the two accept decisions. This is a strong CAI paper and should be evaluated within that context. The authors have also clarified and addressed several major misunderstandings raised during the review process.
While the current evaluation remains limited to a highly controlled experimental setting, future cadaver and clinical studies will be important next steps to further demonstrate the practical and clinical value of the proposed approach. Nevertheless, the work presents sufficient novelty and technical merit to warrant presentation at the meeting, and I believe it will generate significant interest and discussion within the CAI community.
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.
Reviewers appreciated the comprehensive rebuttal, which resolved concerns about the validation on physical 3D phantoms using optical tracking. There was also a strong argument why NeRF and diffusion are unsuitable for this task, given the sparsity of surgical point clouds. While the authors should reduce the claim of clinical readiness (R2) in light of the current evaluation on physical phantom, this level of evaluation is appropriate for a CAI paper, and the paper should generate interest at the conference.
