List of Papers Browse by Subject Areas Author List
Abstract
Spinopelvic measurement (SPM) plays a central role in spinal deformity assessment and surgical planning. Recent AI approaches predominantly adopt multi-stage pipelines that first localize vertebral regions then detect landmarks within isolated patches. While effective and stable, such decomposition restricts models to local subtasks rather than holistic reasoning. As a single-model alternative, direct landmark detection from full radiographs encounters two fundamental challenges: independent landmark prediction suffers from anatomical ambiguity leading to vertebral swapping, and point-wise localization objectives are misaligned with angular measurement goals. We reformulate SPM by replacing point-wise landmark detection with anchor-guided orientation regression that directly models task-relevant geometric primitives. Using frozen DINOv3 vision foundation model encoders with parameter-efficient fine-tuning, we systematically compare three representation methods under controlled experimental conditions—maintaining consistent architectures, decoders, and training protocols to isolate the effect of geometric output formulation. Evaluation on 309 external validation images demonstrates that anchor-guided orientation regression improves measurement consistency (ICC from 0.932 to 0.953) and accuracy (MAE from 2.32° to 2.18°). Critically, the representation maintains ICC above 0.9 even at 256-pixel input where baseline methods degrade significantly. These improvements hold across both ViT and ConvNeXt encoders. This study demonstrates that task-aligned geometric representation—not architectural complexity—is the key determinant of robust, resolution-tolerant automated spinopelvic measurement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2378_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{HwaYun_Beyond_MICCAI2026,
author = { Hwang, Yunseob AND Lee, Youngsuk AND Song, DaeSeok AND Kang, Dong-Ho AND Ro, Du Hyun},
title = { { Beyond Landmarks: Anchor-Guided Orientation Regression for Robust Spinopelvic Measurement } },
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 paper proposes Anchor guided orientation regression for automated spinopelvic parameters measurement. The authors aim to move away from traditional point-based landmark detection, which is prone to vertebral swapping and error propagation. They propose regressing unit orientation vectors anchored to vertebral centers by using a DINOv3 encoder with LoRA fine-tuning. The method was tested on an external dataset of 309 images. The authors claim that this task-aligned representation improves angular accuracy, and maintains performance even when image resolution is significantly reduced.
- 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.This paper is very close to clinical applicability, and this is one of the major strengths of the manuscript. I’m very curious to see how the authors would integrate this method into AP X-rays to detect scoliosis curvature by measuring angles or to detect kyphosis from Thoracolumbar xrays
2.The logic of replacing coordinate regression with orientation primitive regression addresses the mathematical mismatch between the training objective (pixel error) and the clinical goal (angular accuracy).
3.By keeping the encoder and decoder architectures nearly identical across all three tested representations, the authors provide a clean and isolated comparison of the geometric output formulations.
- 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 innovation here is essentially a change in the regression target. Although it’s effective, this is a minor incremental refinement rather than a fundamental advancement.
2.The authors do not provide a baseline using a standard, non-foundation model. It is hard to determine whether the improved angular accuracy is due to “AGOR” formulation itself or simply the result of utilizing a powerful DINOv3
3.The use of a massive model such as DINOv3 remains unjustified.
4.The clinical problem this methodology aims to solve needs to be explained well.
- 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
1.Baseline: To move beyond an incremental claim, the authors must include a comparison using a standard non-foundation encoder such as ResNets trained with the AGOR head. 2.Clinical problem: The manuscript would be stronger if the authors clearly explained how they would integrate this method into a surgical solution or as a diagnostic step.
- 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?
Despite my skepticism regarding the use of foundation models for relatively simple 2D tasks (where real-time inference throughput is important in clinical settings), the authors have demonstrated a sound method that solves a documented clinical problem. Secondly, the task-aligned nature of the AGOR head represents an evidence-driven advancement.
- 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 Authors propose a ConNex or DINO prediction network that predicts heatmaps, offsets points of heatmaps, and angles for a given heatmap. They especially claim that by predicting angles and offsets from a point, they avoid common off-by-one errors and pixel accuracy when only applying points and computing angles afterward.
- 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 paper’s main strength lies in its clinically relevant problem formulation, which addresses automated vertebral assessment in a way that could reduce reliance on handcrafted multi-stage pipelines. A second notable strength is the practical demonstration of feasibility on a comparatively large medical imaging dataset, compared to other papers. The work is also strong in its potential clinical applicability. The ability to directly model vertebral structure and landmarks from imaging data is highly relevant for downstream tasks such as spinal phenotyping, anomaly detection, and large-scale cohort studies. Finally, the paper provides a well-motivated comparison to prior top-down methods, clearly framing the limitations of existing staged pipelines and motivating why a more integrated formulation may be beneficial.
- 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 authors cite prior methods that “adopt top-down pipelines that decompose the task into vertebra localization, region cropping, and landmark detection within isolated regions [5, 7, 9, 18].” However, the subsequent claim that “such decomposition introduces error propagation across stages and restricts models to solving local subtasks rather …” is neither explicitly cited nor empirically validated. To my knowledge, vertebra localization can indeed lead to off-by-one errors, which makes this a plausible concern. Nevertheless, the evaluation would be significantly strengthened by including a direct comparison against such a staged pipeline.
In addition, there is no clear description of how numerical vertebral anomalies are handled. Is the vertebral definition constrained to a fixed number of lumbar vertebrae, or are anatomical variants such as T13/L?, T12/L4, T12/L6, and Castellvi classifications incorporated into the ground truth annotations?
For non-overreaching medical applications, the dataset is reasonably large, but perhaps not sufficiently large to fully justify the use of ViT-based large models. It would therefore be valuable to include experiments with smaller architectures (e.g., ResNet or DenseNet) to assess whether the dataset scale truly warrants larger transformer-based models.
Regarding Table 2, it would likely be more informative to report the training set size rather than the validation set size in this comparison. Including a directly reproduced baseline would also help disentangle the gains attributable to an improved model architecture from those arising from access to a larger dataset. In particular, if an initial step can reliably separate the spine into individual regions, subsequent stages may be less susceptible to local off-by-one errors.
Furthermore, Table 2 reports only the ICC, which is insufficient for a meaningful comparison, as it primarily reflects precision rather than overall agreement or accuracy. Moreover, as far as I understand, the comparison results were not obtained through direct reproduction but were instead taken from the original paper. This weakens and potentially undermines the comparison since the methods were trained and evaluated on different datasets.
- 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 overall contribution is sufficiently interesting. The experimental setup is adequate for demonstrating the approach’s feasibility; however, there is room to clarify the dependencies among the datasets, the model architecture, and the overarching meta-framework. Additional experiments, including reproducing prior work and conducting more comprehensive ablation studies, would further strengthen the paper.
- 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
The paper investigates several variants of solving the target task using a model composed of a frozen DINOv3 encoder, a decoder network, and task-specific prediction heads, with parameter-efficient adaptation via LoRA. Two encoder architectures are considered: ViT-B/16 combined with deconvolutional layers, and ConvNeXt-Base with a feature pyramid network (FPN). While the overall pipeline remains largely consistent, the methods differ in the design of prediction heads (single vs. dual-branch) and in the formulation of the geometric prediction. The results suggest that the choice of geometric representation has a more significant impact on performance than architectural complexity. The authors also report robustness of the proposed method to changes in input resolution.
- 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 paper highlights an important insight that reformulating the geometric representation can lead to notable performance improvements, even when the overall architecture remains largely unchanged. The proposed method demonstrates robustness to reduced input resolution, maintaining a high level of agreement according to the reported metrics. The study is conducted on a relatively large dataset, which supports the reliability of the reported results. Additionally, the authors compare their approach with existing methods, including multi-stage pipelines, in a relevant application setting.
- 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 use of resolution-dependent Gaussian kernels for heatmap generation, together with explicit resizing of input images to target resolutions, suggests that models may have been trained separately for each resolution. In this case, the reported robustness to resolution changes may reflect adaptation to specific scales rather than true generalization across resolutions. Clarification of the training protocol would strengthen this claim. Furthermore, the evaluation primarily relies on aggregated metrics, without providing a detailed analysis of the error distribution. This limits the ability to assess robustness to outliers and variability.
- 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
It would be beneficial to clarify whether models were trained independently for each resolution or using a multi-scale training strategy. This distinction is important for interpreting the reported robustness. Providing additional statistics on error distributions (e.g., standard deviation, median, or percentiles) would offer a more comprehensive evaluation of the method’s performance. Since the SDR metric depends on a predefined threshold (2 mm), presenting performance across multiple thresholds or as a curve would give a more complete picture of the model’s behavior.
- 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 paper presents a meaningful insight regarding the role of geometric representation and demonstrates competitive performance. However, several aspects of the experimental protocol and evaluation could be clarified or expanded to better support the claims, particularly regarding robustness and generalization.
- 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
Dear Area Chair and Reviewers, We sincerely thank the reviewers for their constructive feedback and for recognizing the clinical relevance of our work. The primary goal of our study is to demonstrate that representation reformulation—rather than architectural complexity—is the key determinant of robust angle measurement. 1) Choice of Encoders and Parameter Efficiency (R1, R2): To clarify, we did not utilize massive 7B models. Our framework is built on distilled “Base” variants (ViT-B/16 and ConvNeXt-B) from the DINOv3 paradigm, leveraging their superior pre-trained representations. While the VFM backbones remain frozen, we employ parameter-efficient fine-tuning via LoRA (rank 8, alpha 16) and utilize trainable lightweight decoders. The structural divergence between methods is strictly confined to the task-specific prediction heads and losses (<1%), proving that the performance gains stem from the mathematical alignment of the objective, not architectural scale. 2) Study Scope and Design Rationale (R2): We intentionally focused on standard lateral spinopelvic anatomy (L1–S1 and femoral heads) to strictly validate the technical progress of our AGOR formulation. While prior studies often rely on top-down instance-based ROI modules, our modernized single-model architecture demonstrates superior applicability for focused lateral radiographs. By utilizing anchor-based representations, we achieve stable, class-specific localization without the overhead of additional instance-detection stages. Nevertheless, for whole-spine radiographs involving a large number of vertebrae, we acknowledge that a top-down approach may better align with the clinical counting process and warrants further investigation for such expansion. 3) Resolution Robustness and Evaluation (R1, R4): To ensure a rigorous evaluation, models were trained independently per resolution. We observe that AGOR maintains ICC > 0.9 even at 256px, where baseline methods degrade significantly. This validates that orientation-based primitives maintain superior angular consistency; while spatial distances are highly sensitive to pixel-wise shifts at low resolutions, the underlying directional vectors remain inherently more robust for precise angular measurement. 4) Clinical Significance and Future Directions (R1, R2): While our evaluation followed established metrics in the anatomical landmark detection field, we recognize the need for more rigorous assessment to fully validate clinical utility as suggested. In the final version, we will include a comprehensive Bland-Altman analysis for each parameter to further demonstrate the framework’s reliability in clinical practice. While addressing complex anatomical variations such as L6 or T13 remains a crucial frontier for full clinical translation, we believe our study provides the necessary technical foundation for task-aligned representations in orthopedic assessment.
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 proposes an anchor-guided orientation regression (AGOR) framework for automated spinopelvic parameter measurement. The approach replaces traditional coordinate-based landmark regression with orientation vector prediction anchored at vertebral centers, using a DINOv3 encoder with LoRA adaptation. Reviewers agree that the paper addresses a clinically relevant problem and that aligning the learning objective with angular accuracy is a well-motivated idea (R1, R2, R4). The controlled comparison across geometric representations and the reported robustness to reduced image resolution are also seen as strengths (R1, R4).
However, a central concern across reviewers is the lack of appropriate baselines to isolate the contribution of the AGOR formulation. R1 and R2 both request comparisons with standard non-foundation encoders (e.g., ResNet/DenseNet) using the same prediction head, and question whether the gains stem from the formulation or the use of DINOv3.In addition, R2 notes that claims about limitations of multi-stage pipelines are not empirically validated, despite being a key motivation of the work.
Another recurring issue is the limited depth and rigour of the experimental evaluation. R2 highlights the reliance on ICC as the primary metric and the lack of direct reproducibility in comparisons, while R4 questions whether the reported resolution robustness reflects true generalisation or resolution-specific training. Reviewers also request clearer reporting of training protocols, additional metrics (e.g., error distributions or multi-threshold analysis), and more detailed experimental analysis to support the claims (R2, R4).
Finally, reviewers raise concerns about the clarity of the clinical contribution and the level of novelty. R1 notes that the clinical integration and use case are not sufficiently articulated, and R2 questions how anatomical variability is handled. Both R1 and R4 characterise the contribution as somewhat incremental, focusing on a change in regression target.
Overall, even with the mixed recommendations (R1: Accept; R2: Weak Reject; R4: Weak Accept), I believe most of the concerns can be addressed prior to camera-ready and I recommend that this paper be accepted.