List of Papers Browse by Subject Areas Author List
Abstract
Accurate localisation of anatomical landmarks on distal femur is a critical prerequisite for surgical planning in image-based Total Knee Arthroplasty (TKA). Manual annotation is time-consuming and prone to variability, which can lead to inconsistent implant sizing. Existing classical and deep learning methods remain limited by latency, computational cost, and modality dependence. To address these issues, we operate on segmented 3D surface point clouds and formulate distal femoral landmark localisation as a heatmap regression task that predicts dense probability fields. Ground-truth heatmaps are constructed as Euclidean Gaussians on the femoral surface using learnable landmark specific sigmas. We introduce a hybrid graph-convolution backbone that combines Low-Rank Adaptive Graph Convolution (LR-AGConv) for local feature extraction and Bi-level Routing Point Attention (BRPA) for efficient global semantic modelling. The network is trained using Adaptive Wing (AWing) loss and PointSTAR, a novel point-cloud adaptation of Self-adaptive Ambiguity Reduction (STAR) loss, with uncertainty based multi-task weighting. We curate and release a clinically validated dataset of 550 distal femurs derived from the ShapeMed Knee study with expert landmark annotations. Experimental results demonstrate that the proposed method consistently outperforms state-of-the-art baselines, achieving a Mean Euclidean Error (MEE) of 1.24 ± 0.84 mm across all 11 anatomical landmarks. Code and dataset are available at https://github.com/IGRS-Imaging/SurfMark3D.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6540_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/6540_supp.zip
Link to the Code Repository
https://github.com/IGRS-Imaging/SurfMark3D
Link to the Dataset(s)
https://zenodo.org/records/20711175
BibTex
@InProceedings{B.Dha_SurfMark3D_MICCAI2026,
author = { B., Dharshan AND Hyagreev, Pranav AND Maik, Vivek AND Ansari T. A., Suhail AND Lakshmanan, Manojkumar AND Sivaprakasam, Mohanasankar},
title = { { SurfMark3D: Surface-Based Graph Convolution Framework for Distal Femoral Landmark Localisation in 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 main contributions of the paper are the proposal of a novel point-cloud graph-based landmark localization framework, with memory-efficient and multi-scale graph convolution blocks, in addition to loss functions that integrate surface semantics and uncertainty/multi-task optimization framework.
- 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 major strengths of the paper are:
- Rigid formalization of the proposed graph convolutional blocks and loss functions.
- Comparison of the performance on multiple state-of-the-art and conventional architectures
- Ablation on the constituent terms of the loss functions and convolution blocks
- Showing memory/hardware efficiency in the proposed operators and attention blocks.
- 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 major weaknesses are:
- The paper only tests the framework on cases with KL<=1, which are relatively having smooth surfaces with less complex geometry than higher severity cases. The authors clearly state that the purpose was to ignore the impact of osteophytes; however, it would have been preferable to see the performance of the proposed method on such cases without severe osteophytes.
- The Results section in Page 8 Clinical Generalizability and Cross-modality section suddenly reports experimental results on different datasets and experiment settings that have not been explained before. It would have been better to adequately report those in a different study or expanded version of the paper.
- The results in Table 2 in comparison to previous methods is minimal, where it’s almost in submillieter Euclidean error level. Given that all cases are KL<=1, the improvement level may be clinically insignificant, with the major improvement in the computational efficiency side.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
Despite the clear derivations and innovation in the model architecture, loss function and optimization framework, which has been reflected on more efficient computations, the clinical significance of the results is unclear. This led me to a weak accept decision.
- 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?
TKA, statistical shape models, surface analysis, landmark localization
- 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.
In my opinion, the authors have properly addressed my and other reviewers’ points in the rebuttal phase. I’m encouraged to accept the paper upon the rebuttal response.
Review #2
- Please describe the contribution of the paper
This paper presents a surface-based hybrid graph convolution framework for automatically localising distal femoral landmarks in Total Knee Arthroplasty. Instead of relying on volumetric data, the method works directly with 3D point clouds and predicts probability heatmaps for each landmark with learnable uncertainty. The proposed architecture combines Low-Rank Adaptive Graph Convolution to capture local geometric features with Bi-Level Routing Point Attention to incorporate global contextual information. In addition, the authors introduce PointSTAR, a loss function designed to account for anisotropic uncertainty on anatomical surfaces. Experimental results on a dataset of 550 femurs show improved performance over several baselines, with a reported mean error of 1.24 ± 0.84 mm, approaching inter-observer variability.
- 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 demonstrates strong clinical relevance for surgical planning in Total Knee Arthroplasty. The learnable Gaussian spreads and the proposed PointSTAR loss effectively address anisotropic ambiguity and are well-motivated for this task. The whole framework is technically reasonable and clear. The use of multiple evaluation metrics provides a comprehensive assessment of performance. The release of a clinically validated dataset with expert annotations is valuable for the community and enhances reproducibility.
- 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.
Although the proposed workflow address the problem, the improvements over baselines are relatively small, and no statistical significance testing is provided. The framework includes multiple components (LR-AGConv, BRPA, AWing, PointSTAR, uncertainty weighting). Although ablations are provided, it remains unclear which components are essential vs. marginal contributors to performance gains. Some aspects are not clearly described in the paper. For example, it would be helpful to define MEE in the abstract. In addition, SDR is used throughout the paper without any explicit explanation.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 proposed method is clinically relevant and technically well-founded, with solid experimental support including ablation studies. However, the overall performance gains are relatively modest, and no statistical significance analysis is provided to support these improvements. While the paper is generally clear, the presentation could be improved by more consistently defining and organizing abbreviations.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed my main concerns. In particular, they provided statistical analysis using Wilcoxon signed-rank tests and clarified that the proposed method achieves statistically significant improvements over most baselines, while acknowledging that some differences are not significant due to performance approaching the inter-observer variability floor. They also agreed to define MEE and SDR more clearly in the final version. I consider the paper’s technical contribution, reproducibility, and dataset release sufficient for acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes a modality-agnostic framework for automated distal femoral landmark localisation operating on segmented 3D surface point clouds. The key technical contributions include a hybrid backbone combining Low-Rank Adaptive Graph Convolution (LR-AGConv) for memory-efficient local feature extraction and Bi-Level Routing Point Attention (BRPA) for selective global context modelling, a novel PointSTAR loss that decomposes anisotropic uncertainty along surface normal and tangent directions, learnable landmark-specific Gaussian sigmas for heatmap regression, and the release of a 550-case expert-annotated distal femur benchmark dataset. The method achieves 1.24±0.84mm mean Euclidean error across 11 landmarks.
- 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.
Public benchmark dataset. The release of 550 expert-annotated distal femur meshes with inter/intra-observer variability reporting (1.24±1.05mm and 1.18±0.90mm respectively) fills a genuine gap in orthopedic landmark research, where comparable public datasets are essentially nonexistent. This alone provides lasting value to the community regardless of the proposed method’s merits.
Memory-efficient engineering. The low-rank decomposition of AGConv achieves 43.1% training memory reduction (40.4GB→23.0GB) without accuracy degradation, and the full pipeline requires only 1GB VRAM at inference with 1.7s latency. These are practical engineering contributions with applicability beyond this specific task.
PointSTAR loss formulation. The decomposition of prediction uncertainty into surface-normal and tangent-plane components is geometrically well-motivated for surface-constrained landmarks. The ablation demonstrates concrete improvements for geometrically ambiguous landmarks such as FLTA (1.64→1.41mm) and FMCD (1.86→1.60mm), and accelerates training convergence from >100 to ~60 epochs.
- 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.
Absence of volumetric end-to-end baselines. The paper claims advantages over volumetric and multi-view methods [2,14] but provides no direct comparison. Berger et al. [2] demonstrated that nnU-Net operating directly on CT volumes achieves accuracy statistically equivalent to manual annotation for 6 distal femoral landmarks without requiring any upstream segmentation. The paper’s Table 2 compares only point cloud backbones (PointNet++, PointMLP, PTv3), proving superiority within the point cloud paradigm but failing to justify why the point cloud paradigm itself is preferable to established volumetric approaches. This is a fundamental gap in experimental design.
Severe clinical population mismatch. The dataset is restricted to KL≤1 (non-osteophytic) subjects, which essentially excludes the actual TKA patient population (predominantly KL 3-4). TKA candidates present with osteophytes, subchondral sclerosis, bone cysts, and varus/valgus deformity that drastically alter surface geometry. The 1.24mm MEE reported on near-normal femora cannot be extrapolated to pathological anatomy, and no analysis is provided on how model performance would degrade with increasing disease severity. This critically undermines the paper’s claim of clinical relevance for TKA planning.
Misalignment between task definition and clinical workflow. In actual TKA surgical planning, individual landmark localisation is not the clinical endpoint. Sizing is determined by implant template fitting after cutting plane simulation, where ML width and AP dimension are evaluated simultaneously relative to the implant catalog. Notching assessment emerges directly from anterior resection plane simulation, not from individual landmark coordinates. The paper reports no downstream surgical planning metrics (e.g., implant size selection accuracy, component alignment error, or cutting plane deviation), making it impossible to assess clinical impact.
Unquantified error propagation from upstream segmentation. The pipeline requires MRI→segmentation→mesh→FPS before landmark detection begins, yet no analysis of how upstream segmentation errors propagate to landmark accuracy is provided. The Laplacian smoothing (15-25 iterations) applied to meshes further introduces uncharacterised geometric distortion. The claimed “modality-agnostic” advantage is undermined by this dependence on accurate segmentation, which itself is modality-specific.
- 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.
- 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
The cross-modality validation on 56 CT-derived VSD point clouds (1.5±0.9mm) is acknowledged but insufficient — the KL grade distribution and osteophyte status of these cases are not reported, precluding assessment of generalisability to pathological anatomy. The scapula experiment demonstrates architectural flexibility but is irrelevant to the claimed TKA application. The authors should consider reframing their contribution as a general-purpose point cloud landmark detection method rather than a TKA-specific clinical tool, or alternatively, provide comprehensive validation on KL 2-4 subjects with downstream surgical planning metrics.
- 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 engineering contributions (LR-AGConv memory efficiency, PointSTAR loss, public dataset) are recognised, and the ablation study is methodologically sound. However, the absence of volumetric end-to-end baselines leaves the central methodological claim unjustified, and the restriction to KL≤1 subjects creates a fundamental disconnect with the stated TKA application. The task formulation itself — individual landmark localisation rather than surgical plan optimisation — reflects a gap between the computer vision problem definition and actual clinical needs. These are not issues addressable through minor revisions.
- 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
Author Feedback
We thank reviewers for recognising our dataset (“filling a genuine gap”, R3), architectural contributions “strong clinical relevance” (R2), “rigid formalization” (R1), and “geometrically well-motivated” formulation (R3). Main concerns addressed below. [A] KL<=1 & scope reframing (R3, AC): We adopt R3’s suggestion: this is a general-purpose, modality-agnostic surface landmark framework, with TKA as the primary validated application. Zero-shot transfer to a distinct anatomy (scapula, 1.8+/-1.35 mm) supports this. Selecting KL<=1 is a ground-truth validity requirement, not an arbitrary restriction. The 11 landmarks are distinct morphological features whose definitions break down on osteophytic surfaces, causing annotation ambiguity. Horteur [8] demonstrates that degenerative morphological changes (osteophytes, bone wear) substantially increase both manual and automatic landmarking errors underscoring the difficulty of establishing reliable ground truth on advanced OA surfaces. Grammens [5] and Berger [2] similarly exclude/defer these cases. Clinically, osteophyte handling is a dedicated upstream preprocessing step prior to landmarking and recent works also treat virtual osteophyte removal as a separate module. [B] Volumetric baselines & cross-modality (R1, R3, AC): While we acknowledge the strength of volumetric baselines like nnU-Net, a direct empirical comparison is challenging due to fundamental differences in problem formulation. Our goal (Sec. 1) is modality-agnostic surface processing. Volumetric methods are inherently modality-dependent: Berger’s [2] nnU-Net consumes raw CT volumes, rendering it inapplicable to MRI. Conversely, our framework accepts segmented meshes from any modality: a single surface model covers MRI and CT-derived meshes without retraining, which volumetric methods cannot structurally offer. Furthermore, they report a higher MEE (2-3mm vs our 1.24mm). Our VSD zero-shot result (1.5+/-0.9mm on CT-derived meshes) bridges this gap: our model achieves near-manual accuracy on CT derived meshes (without CT-specific retraining) empirically validating our paradigm. [C] Downstream planning metrics (R3, AC): We agree downstream integration strengthens clinical impact. Accurate localisation is the foundational prerequisite for downstream measurements; implant sizing, alignment, and cutting-plane definitions derive from these coordinates. Downstream integration remains future work (Sec. 4). We will clarify this scope in the camera-ready. [D] Robustness to upstream variation (R3, AC): ShapeMed [4] reports femoral DSC 0.99/ASSD 0.08-0.15mm, establishing expert-level meshes. Our pipeline (segmentation, Laplacian smoothing, FPS downsampling) precedes model input. Our results demonstrate robustness across these stages. Crucially, the VSD [9] cohort (56 CT-derived meshes, including OA-affected) uses an entirely independent upstream pipeline, confirming robustness to simultaneous shifts in modality, upstream pipelines, and smoothing. [E] Statistical evidence & components (R1, R2, R3, AC): As requested, we applied Wilcoxon signed-rank tests to our existing Table 2 predictions. Our framework achieves statistically significant improvements (p<0.05) over the majority of baselines (PTv3, PointNet++, PointMLP, DeltaConv, PTv2/v3 Attn). For the remaining baselines (PAConv, Self Attn) and intra-ablation variants, differences do not reach p<0.05.This is mathematically expected: because all methods are converging near the human inter-observer floor (1.24+/-1.05 mm) where even our base model already operates (1.28+/-0.90 mm) MEE reduction naturally plateaus. Instead, as R3 appreciated, our meaningful improvements lie along orthogonal axes: architectural efficiency, convergence, and hard-landmark robustness. These improvements hold consistently across all metrics, all 11 individual landmarks (Fig. 3), and three independent runs. [F] Abbreviations (R2): MEE and SDR will be explicitly defined in the camera-ready.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a point‑cloud / surface based graph framework to localise distal femoral landmarks for Total Knee Arthroplasty planning. It works on segmented 3D surface point clouds and predicts landmark heatmaps. A 550‑case expert‑annotated dataset is released as well.
The clinical population does not match the target. The dataset includes only KL ≤ 1.This means almost normal anatomy. TKA patients usually have KL 3–4 with osteophytes and deformity. These changes surface geometry a lot. So the reported ~1.2 mm errors on smooth femora may not transfer to real TKA planning. The paper does not test severe cases or discuss how performance changes.
Important baselines are missed. It compares mostly the point‑cloud frameworks. It does not compare against strong volumetric end‑to‑end methods (e.g., nnU‑Net on CT volumes) that can localise landmarks without upstream segmentation and meshing. Without this, the paper does not justify why point clouds are better for this task.
Reviewers were also commenting a mismatch of th workflow related to clinical endpoints. In practice, surgeons care about implant sizing, alignment, and cutting planes. They do not use landmark coordinates alone as the final output. The paper does not report downstream planning metrics (implant size accuracy, alignment errors, plane deviation). This weakens the clinical impact claim.
The pipeline depends on upstream segmentation and mesh processing. Segmentation errors can propagate to landmark errors. The paper does not quantify this. Laplacian smoothing and sampling may also distort geometry, but the paper does not analyse this.
Statistical evidence is weak. The paper reports small improvements over baselines, but it does not report statistical significance.
Reviewers also noted unclear reporting of abbreviations (e.g., MEE, SDR) and some results appear suddenly with limited explanation.
Overall, the paper has good engineering contributions and a valuable dataset release. But the current validation does not support strong clinical claims for TKA planning. The missing volumetric baselines and the KL≤1 restriction are the biggest gaps.
- 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 rebuttal clarifies some points. The statistical testing and dataset release are positive. However, they do not overcome the main concerns.
The authors explain that KL≤1 cases were selected for reliable landmark annotation. But it does not solve the clinical mismatch (the paper is framed around TKA planning). It only explains why the dataset is easier and more reliable to annotate.
Second, the modality-agnostic claim: the method is only modality-agnostic after segmentation. The full pipeline still requires CT or MRI segmentation before mesh generation. The paper is not sufficiently validated as a TKA planning method.
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.
Majority of reviewers vote for acceptance. There are a number of critical concerns regarding methodology and experimental validation, however, I think these have been addressed reasonably well. The dataset that will be published is acknowledged and a big pro for this work. With a reformulation of the contribution regarding the intention of the work and a noting of the limitation that TKA disease cases are not the target of this initial work, the manuscript could be of interest to the MICCAI 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.
The authors have effectively addressed the reviewers’ concerns, particularly those raised in the meta-review. I find the responses convincing, including those regarding the validation choices, volumetric baselines, and statistical significance analyses. Overall, I believe this is a solid paper, and I am inclined to support its acceptance.