List of Papers Browse by Subject Areas Author List
Abstract
Accurate spine segmentation is pivotal for surgical planning, yet remains hindered by complex anatomical structures and ambiguous boundaries. Recently, Segment Anything Model (SAM)-based methods have shown impressive generalization. However, single-level prompts often fail to fully capture semantic discriminability and anatomical consistency in spine imaging. To address these challenges, in this paper, we propose an anatomy-aware hierarchical prompting refinement network for spine image segmentation, called AHPR-Net. Specifically, to enhance the semantic richness of prompt information, we propose a hierarchical multi-level prompting strategy that jointly incorporates categorical semantics, anatomical relational priors, and morphological shape information to provide anatomically grounded guidance for spine segmentation. Then, to effectively fuse multi-level prompt information, we design a dynamic prompt tuning module that adaptively selects and enhances the most informative prompts and improves the perception ability of the complex anatomical structures of the spine. Furthermore, richer anatomical prompts alone cannot eliminate boundary ambiguity and the resulting uncertainty, we propose an uncertainty-guided refinement strategy that explicitly models boundary uncertainty to improve structural reliability during training. Experimental results on an in-house spine dataset and a public SPINDER dataset demonstrate that our proposed method achieves superior segmentation performance compared to several state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3186_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zjy399/AHPR-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaJun_AHPRNet_MICCAI2026,
author = { Zhao, Junyong AND Wang, Kun AND Fan, Dingwei AND Dou, Qi AND Sun, Liang AND Zhang, Daoqiang},
title = { { AHPR-Net: Anatomy-aware Hierarchical Prompting Refinement Network for Spine Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes AHPR-Net, a novel framework for spine image segmentation that addresses the limitations of single-level cues based on the SAM model. Its main contributions are threefold: (1) a hierarchical multi-level cues strategy that integrates categorical semantics via CLIP, anatomical relation priors via Llama, and morphological shape information from multi-atlas data (AI); (2) a dynamic cues tuning module that adaptively weights and fuses the embeddings of the three cues based on the context of the input image; and (3) a guided refinement strategy that explicitly models the predicted distribution and utilizes Monte Carlo sampling to mitigate boundary blurring. This method achieves state-of-the-art performance on both the internal CT dataset and the publicly available SPIDER 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 prompt design in this paper is innovative and intuitive. Traditional SAMs typically use only points or boxes as cues, which is insufficient for complex spine images. The authors introduce a Large Language Model (Llama) to handle spatial topological relationships and use multiple graphs to provide shape priors. Applying the Llama approach to CV processing is a very good cross-modal and multimodal approach.
(2) The module fusion design is ingenious. For the three different prompts mentioned in the paper, the authors use a Dynamic Prompt Tuning mechanism, allowing the model to calculate weights based on global image features, which is much better than directly concatenating all features. Furthermore, to address the issue of blurred boundaries in medical images mentioned in the paper, the authors designed an “Uncertainty-guided Refinement” module.
- 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) Computational Cost. The paper uses Monte Carlo sampling to estimate model uncertainty, but this is often very time-consuming. However, the paper makes no mention of the computational cost, which raises questions about its practical application in scenarios requiring high frame rates.
(2) Robustness of Pre-registration. The shape-level prompt in the paper relies on “pre-registration multi-atlas” of spinal data. Non-rigid registration of medical images is inherently challenging. In cases such as scoliosis or severe fracture deformities, registration is highly likely to fail. Furthermore, the authors do not discuss whether the model has the ability to correct errors if the registration provides incorrect priors.
Personal Questions: The paper mentions the concept of “priors. “ Does this method place more trust in model predictions or observations? Are there parameters similar to “Kalman gain”? If so, how can their value be justified? If not, how are observation noise and model noise adjusted to affect the results? How can the rationality and effectiveness of this adjustment mechanism be explained?
Madesta, Frederic, et al. “Monte Carlo-based simulation of virtual 3 and 4-dimensional cone-beam computed tomography from computed tomography images: An end-to-end framework and a deep learning-based speedup strategy. “ Physics and Imaging in Radiation Oncology 32 (2024): 100644.
Zhao, Lei, et al. “SpineRegNet: Spine Registration Network for volumetric MR and CT image by the joint estimation of an affine-elastic deformation field. “ Medical Image Analysis 86 (2023): 102786.
- 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 most significant highlight of this paper is the deep integration of multi-layered language and shape cues. Its combination of dynamic cue adjustment and uncertainty-aware refinement directly addresses a core challenge in spine segmentation.
Despite some computational overhead in Monte Carlo sampling and issues with pre-registration, the method’s innovation and robust experimental results demonstrate its excellence.
- 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
This paper proposes a SAM-based spine image segmentation network named AHPR-Net. By integrating categorical semantics (CLIP), anatomical spatial relationships (Llama), and multi-atlas shape priors, the network constructs a hierarchical prompting mechanism and employs dynamic routing weights for fusion. Furthermore, to optimize complex boundaries, an uncertainty refinement module based on mean-variance estimation and Monte Carlo sampling is introduced at the output stage. This method achieves performance superior to current state-of-the-art (SOTA) models (e.g., nnU-Net, SAM_Med2D, and U-Mamba) on both an in-house CT dataset and the public SPIDER MRI 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.This paper transcends the limitations of traditional SAM, which relies solely on point or box prompts. The proposed method ingeniously integrates high-level clinical semantics and anatomical topological relationships extracted by large language models (LLMs) with traditional atlas shape priors, thereby forming a structurally complete guidance. 2.The method proposed in this paper was validated on dual cross-modality datasets (CT and T1-MRI) against highly comprehensive baselines, covering the latest Transformer, Mamba, and SAM variants, demonstrating its strong generalization capabilities. 3.The ablation studies in this paper are comprehensive in scope. Specifically, the experiments not only independently evaluate the effectiveness of the LoRA fine-tuning and the uncertainty-guided refinement module , but also conduct an in-depth comparison of various prompt fusion strategies, such as concatenation, summation, and the proposed dynamic fusion. Furthermore, this paper performs systematic combinatorial ablation tests on the hierarchical prompting mechanism, including class-level (CLS), relational-level (REL), and shape-level (SHA) prompts , while also exploring the sensitivity of the key hyper-parameter regarding the number of atlases.
- 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 dynamic fusion mechanism in the paper is overly simplistic, relying solely on simple Global Average Pooling (GAP) and Softmax for weight allocation (Equation 4). For highly heterogeneous features (textual semantics vs. spatial topology vs. morphology), such a linear weighting strategy can easily lead to feature homogenization. 2.The paper’s current “Class-Level Prompt” relies on highly templated, hard-coded text (e.g., “This CT image includes…”). Such fixed templates restrict the upper bound of the multimodal model. Directly extracting and utilizing clinical text reports, or leveraging LLM to generate pathological descriptions , would be more effective. 3.The ablation study design still contains a critical gap: the lack of cross-combinatorial validation among the three core contributions. Specifically, the authors should provide a set of progressive ablation experiments (e.g., Baseline vs. Baseline+Prompts vs. Baseline+DynamicTuning vs. Baseline+Prompts+DynamicTuning vs. Full Model) to explicitly quantify the incremental and cumulative contributions of each module within the overall framework.
- 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 core premise of this paper is clear and the proposed method achieves excellent performance in the clinically challenging task of spine segmentation. While the breadth of the hierarchical prompt design is commendable , there remains significant room for further exploration regarding the deep fusion of heterogeneous features.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper presents AHPR-Net, which incorporates three types of prompts - category semantics, anatomical relationships, and shape priors - into a SAM-based spinal segmentation framework, and combines dynamic prompt fusion and uncertainty refinement. It has achieved good results on a self-built CT dataset and the publicly available SPIDER dataset. The overall topic selection is clinically significant, the method framework is relatively complete, and the experimental coverage is extensive, including mainstream CNN, Transformer, Mamba, and SAM-based methods.
- 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 issue is significant and the application is clear: The multi-structure segmentation of the spine has practical value for preoperative planning, and the objective of the paper is clear. 2.The method combination is relatively complete: multi-level prompts, dynamic prompt tuning, and uncertainty refinement form a relatively coherent framework. 3.The experimental results are competitive: they outperform the multi-class baselines on both datasets, especially with a significant improvement in the 95HD metric. 4.The ablation experiments basically cover the core modules: the effects of LoRA, uncertainty refinement, prompt fusion methods, multi-level prompts, and the number of atlases have been verified, meeting the basic requirements within the scope of a conference paper.
- 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 authors are encouraged to reduce broad and generic statements such as “uncertainty” and “LLM medical knowledge,” and instead focus the core motivation on the main issue: single-level prompting lacks anatomy-aware hierarchical guidance for spine segmentation. 2.The method details still need to be expressed more precisely. In Eq. (4), the definition of 𝐹𝑖𝑛 is inconsistent: in the preceding text it denotes image features, while in the equation it becomes the concatenation of prompts. It is also unclear whether the dynamic gating depends on image context. Within the limited space, the authors are advised to prioritize clarifying the symbols and data flow, otherwise the method may remain difficult to understand. 3.The practical role of the prompts needs stronger justification. The category prompts and anatomical relation prompts appear to be fixed text. If all samples share the same prompts, their information gain may be limited. The authors are advised to briefly clarify whether the prompts are fixed, whether they are designed separately for CT and MRI, and how their embeddings are aligned with image features. 4.There are concerns regarding the uncertainty module equations. In Eq. (7), the sampling form is written as 𝜇+𝜖𝜎2, whereas it is usually 𝜇+𝜖𝜎. If this is an intentional design, it should be explained. Another issue is that 𝑈𝐺𝑇 is not used in Eq. (8), but it is used in the loss, making the description of the pipeline seem somewhat abrupt. 5.The description of experimental fairness is insufficient. Since many comparison methods are included, the authors are advised to add a sentence clarifying whether all baselines were re-trained under the same data split, input size, and training protocol. For SAM-based methods, it should also be specified whether the same prompt settings were used across all methods. 6.“SPINDER” and “SPIDER” are spelled inconsistently. 7.“AHPR-Net(Our)” is recommended to be changed to “AHPR-Net (Ours)”. 8.In Table 2, the IoU improvement brought by a single prompt is not stable. The authors are advised to briefly explain this in the main text.
- 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.
(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 overall direction of this work is reasonable, the results are strong, and the ablation studies are basically sufficient. The main weakness lies in the presentation of the method and the insufficiently solid justification of prompt effectiveness. If the authors can clarify the key equations, provide lightweight validation of the role of prompts, and strengthen the description of experimental fairness in the rebuttal or final version, the paper would be acceptable.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank all reviewers for their valuable comments and constructive suggestions. We appreciate the encouraging comments like “The prompt design in this paper is innovative and intuitive” of R1, “This paper transcends the limitations of traditional SAM” of R2, and “The issue is significant, and the application is clear” of R3.[R1] Q1 Computational Cost: We thank the reviewer for raising this important concern regarding computational efficiency. We agree that Monte Carlo (MC) sampling may introduce additional inference overhead compared with deterministic inference. In the future, we will include the runtime analysis. We also appreciate the reviewer’s suggestion regarding acceleration strategies. [R2] Q2 Multi-atlas pre-registration: We appreciate the reviewer’s insightful comment regarding the robustness of pre-registration. We would like to clarify that the pre-registration atlas is a coarse anatomical prior to provide high-level topology guidance. The final prediction is still primarily determined by the image-driven segmentation branch. We agree that severe deformities such as scoliosis or fracture may challenge registration-based priors. In the future, we will investigate the impact of complex structures and propose reasonable solutions. [R2,R3]Q3 Cross-combinatorial ablation study: We thank the reviewer for this important suggestion. We agree that progressive cross-combinational ablation studies would better demonstrate the individual and cumulative contributions of each component. In the future, we will add additional ablation experiments. [R3]Q4 Writing issues: In our final revision, we will carefully revise the paper and correct the issues with spelling inconsistencies, repeated abbreviation definitions, and standardize the mathematical notation.
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 AHPR-Net, a SAM-based framework for spine segmentation that integrates hierarchical multi-level prompts, including categorical semantics (via CLIP), anatomical relations (via an LLM), and multi-atlas shape priors, alongside dynamic prompt fusion and uncertainty-guided refinement. Reviewers agree that the paper addresses a clinically relevant and technically challenging problem, and that the integration of multi-level cues within a unified framework represents a meaningful advance over conventional SAM prompting strategies (R1–R3). The reported performance across both CT and MRI datasets is strong, and the breadth of comparisons is generally viewed as comprehensive (R2, R3).
A key strength highlighted across reviews is the effective combination of heterogeneous information sources. In particular, the use of LLM-derived anatomical relationships and atlas-based priors to complement visual features is seen as a promising cross-modal direction (R1, R2), while the dynamic prompt tuning mechanism is considered more principled than simple feature concatenation. The uncertainty-guided refinement module is also viewed positively for improving boundary quality (R1, R3), and the ablation studies are broadly adequate in assessing individual components such as prompt types and fusion strategies (R2, R3).
However, several concerns remain. The design of certain components appears relatively simple or insufficiently justified—for instance, the GAP + softmax-based fusion may be too limited for heterogeneous features (R2), and the actual contribution of prompts is unclear if they rely on fixed templates (R2, R3). There are also issues with methodological clarity, including inconsistent notation and unclear formulations (R3), as well as gaps in experimental reporting, such as missing progressive ablations and limited discussion of fairness across baselines (R2, R3). Practical considerations, including computational overhead from Monte Carlo sampling and robustness of atlas registration in pathological cases, are also underexplored (R1). Most of these concerns are likely to be addressable prior to the camera-ready version. Overall, the work is viewed positively by all three reviewers, and I recommend provisional acceptance.