List of Papers Browse by Subject Areas Author List
Abstract
The deployment of PET tracers faces a cold-start dilemma in two common scenarios: when a novel tracer enters early clinical use with scarce healthy controls (HCs) and no diagnostic labels, and when an established tracer is introduced at a new site or resource-limited region where normative references have yet to be accumulated.
Existing anomaly detection methods require large training cohorts or operate only on two-dimensional slices, failing to address the joint demands of extreme data scarcity and volumetric reasoning.
We propose \textbf{TracerAD}, a training-free few-shot framework for three-dimensional brain PET anomaly detection that requires only a handful of HC scans and no model training.
TracerAD extracts frozen vision foundation model features from registered brain volumes and detects pathological deviations at three spatial scales: Spatially-Anchored Focal Detection (SAFD), Region-Constrained Semantic Matching (RCSM), and Topology-Aware Metabolic Reasoning (TAMR).
This multi-scale approach produces a clinically interpretable 3D anomaly heatmap.
Evaluated on ADNI amyloid PET under 1- to 10-shot settings, TracerAD achieves 84.13\% AUC at 10-shot, with heatmaps aligning to known amyloid deposition patterns—offering radiologists an immediately deployable tool for cold-start scenarios.
Our code will be released at \href{https://github.com/MedAIerHHL/TracerAD}{GitHub}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3396_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{HuaHao_TracerAD_MICCAI2026,
author = { Huang, Haolin AND Wu, Junlei AND Lu, Jiaying AND Shen, Zhenrong AND Wang, Xinyu AND Zuo, Chuantao AND Wang, Qian},
title = { { TracerAD: Training-Free Few-Shot 3D Anomaly Detection for Novel PET Tracers } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of this paper is to formulate the PET cold-start setting where only a handful of healthy controls are available for a novel tracer or a new site as a training-free 3D few-shot anomaly detection problem. To address it, the authors propose TracerAD, which extracts frozen VFM features from registered brain PET volumes and combines SAFD for coordinate-level focal matching, RCSM for atlas-constrained within-region matching, and TAMR for graph-based inter-regional metabolic reasoning, producing an interpretable 3D anomaly heatmap without model training.
- 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.
First, the problem setting is highly clinically relevant. The paper directly targets a realistic PET deployment scenario in which only a few healthy references exist and no patient labels are yet available, which is more practical than standard large-cohort supervised training. Second, the method is well aligned with clinical intuition. SAFD captures fine focal deviations, RCSM adds robustness to within-ROI variation and registration imprecision, and TAMR models circuit-level metabolic disruption, making the framework more compelling than a single-scale retrieval scheme. Third, the reported empirical results are promising. At 10-shot, TracerAD reaches 84.13% AUC, 78.83% weighted F1, and 81.60% sensitivity, outperforming the supervised baseline in most settings, and it achieves 86.60% sensitivity at 2-shot. The voxel-wise 3D heatmaps are also a meaningful strength because they add anatomical interpretability beyond subject-level labels. Fourth, the ablation study is reasonably informative. The paper shows additive gains from SAFD to RCSM to TAMR and suggests that sparse graph reasoning improves both performance and stability.
- 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 biggest weakness is that the algorithmic novelty is somewhat narrower than the framing suggests. SAFD is closely related to patch-matching and retrieval-based anomaly detection ideas such as Roth et al. ’s Towards Total Recall in Industrial AD (CVPR 2022) and RAD (Long et al. , 2026), while RCSM and TAMR are explicitly close in spirit to UniVAD’s component-aware matching and graph-based interaction module (Gu et al. , CVPR 2025). In that sense, the paper’s novelty lies more in problem formulation and PET-specific integration than in a fundamentally new core formulation. Second, the comparison is limited. The main baseline is a supervised few-shot classifier from Liu et al. , so it remains unclear how TracerAD compares against stronger unsupervised or training-free anomaly detection baselines that are methodologically closer to the proposed approach, especially retrieval/matching-based ones such as UniVAD or RAD. Third, The 3D heatmaps are compelling qualitatively, but there is no voxel-/region-level ground-truth evaluation or reader study to establish localization validity. Fourth, the framework relies heavily on registration and atlas quality. Small misregistrations may affect SAFD and RCSM substantially, yet no robustness analysis is provided. The equal-weight fusion is also reasonable but somewhat heuristic.
- 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?
I lean slightly positive because the paper tackles a clinically important low-data PET deployment problem and presents a practical, interpretable, and reasonably strong training-free 3D solution. However, the core method is largely an adaptation/integration of prior retrieval and graph-based ideas, and the experimental scope is limited to one tracer with a narrow set of baselines. Overall, I view it as marginally above the acceptance threshold because of the importance of the problem and the practical value of the proposed framework.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes TracerAD, which is a training-free few shot framework for 3D brain PET anomaly detection in cold-start tracer deployment scenarios. The method uses frozen VFM features extracted from PET volumes and combines it, which are SAFD, RCSM, TAMR. The output are in voxel wise 3D anomaly heatmap. The method is evaluated on ADNI amyloid PET under 1,2,4,8,10 shot protocols using only healthy controls as references without model training or fine-tuning.
- 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 shows clear and practical relevance of clinical problem. 1.The method is using training-free few shot setup which can be useful in the case of labeled patient data is unavailable.
2.The method shows competitive performance compared to supervised learning baselines using stricter unsupervised protocol.
- 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 paper emphasizes voxel wise 3D anomaly heatmaps, but the evaluation is only showing AUC, weighted F1 score, and sensitivity. It would be better with voxel level, lesion localization metrics to verify that the heatmap is accurate.
2.The visual evidence is not very convincing. In Fig2, it only shows average heatmaps across all tests, but it would be more convincing with case level examples with meaningful abnormal regions.
- 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
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 paper presents with a well motivated problem and proposes moderate novelty of training-free framework. The validation is limited to a single dataset, and the localization benefit is not quantitatively verified.
- 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 authors model the cold-start problem prominent in PET diagnoses as an anomaly detection problem and propose TracerAD to tackle it. They have taken into consideration the lack of availability of data and reference priors, and opt for a training-free approach for practical application in clinical settings. The main idea is that it works in a fine-to-coarse approach across three spatial scales to address the focal variations, region-based patterns and inter-regional relationships.
- 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 is well written and the authors have explained the ideation and importance of each of the individual branches (SAFD, RCSM, TAMR) in anomaly detection very well in terms of formulation and as well as ablation studies. The visualisation in the overview diagram is also well-illustrated.
- 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.
- While the proposed framework shows a promising performance in terms of experiments and results, the authors have not provided implementation details and inference time analysis which makes it difficult to assess reproducibility and real-world applications.
- The comparison with other self-supervised approaches is limited. It would be helpful to include comparisons with backbones from other paradigms, rather than focusing only on various DINO-based models.
- 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
There is a typo in the title of Table 2 (“backbone” is misspelled). To improve readibility, the authors may consider including abbreviations of MNI and AAL.
- 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?
The paper is well written and logically structured that makes it readable for both experts and readers with a background in medical image analysis. Due to its strong conceptual clarity and sound technical formulation, this paper makes a meaningful contribution to the field and warrants acceptance.
- 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
Thanks for the valuable comments. We appreciate the recognition of our work’s clinical relevance and have addressed each concern as follows. 1.Method design and positioning (R1): We acknowledge the intellectual lineage from retrieval-based anomaly detection. TracerAD is not a direct adaptation but a domain-specific redesign for 3D brain PET. Existing methods such as UniVAD rely on component segmentation modules designed for industrial images, where each image contains a small number of distinct objects. Brain PET volumes, by contrast, encompass over a hundred fine-grained anatomical regions with continuous metabolic gradients, making learnable component segmentation inappropriate. TracerAD instead leverages the AAL3 atlas, a widely adopted neuroanatomical parcellation, to define matching regions directly. Similarly, SAFD exploits the coordinate-level alignment provided by MNI152 registration, which is more precise than the local neighborhood matching used in 2D retrieval methods. To our knowledge, TracerAD is the first training-free few-shot framework for 3D brain anomaly detection. 2.Baseline comparison (R1): The supervised baseline from Liu et al. uses the same DINO backbones, providing a fair comparison under a consistent feature extraction pipeline. The key finding is that TracerAD, under a stricter unsupervised protocol using only N unlabeled HCs, matches or exceeds a supervised method requiring 2N labeled samples. Existing training-free methods such as UniVAD and RAD operate on 2D natural images and are not directly applicable to 3D brain PET without substantial redesign. 3.Heatmap localization evaluation (R1, R2): Unlike tumors with well-defined boundaries, PET tracer deposition in neurodegenerative diseases is diffuse and spatially heterogeneous, and no pixel-level ground-truth annotations exist for such patterns. In clinical practice, nuclear medicine physicians describe regional uptake qualitatively rather than delineating precise lesion boundaries, even for established tracers. For novel tracers, this challenge is further compounded by the absence of known deposition patterns. Voxel-level localization metrics such as Dice score are therefore inapplicable, and subject-level classification metrics remain the standard evaluation protocol in PET anomaly detection. We note that the group-average heatmaps in Fig. 2 show patterns consistent with known amyloid deposition, providing preliminary evidence of localization validity. Individual case visualization will be provided in the journal extension due to the conference page limit. 4.Registration robustness (R1): MNI152 registration and AAL3 parcellation are the standard pipeline in neuroimaging research. The RCSM branch addresses registration imprecision through region-level average pooling before matching, absorbing minor spatial shifts within each atlas region. We will explore personalized atlas segmentation and registration-free variants in the journal extension. 5.Implementation details and inference time (R3): We will release the source code upon acceptance to ensure full reproducibility. Unlike routine disease screening, our cold-start scenario prioritizes diagnostic accuracy over inference speed, as manually interpreting scans from unfamiliar tracers is extremely challenging. 6.Backbone and application scope (R3): We selected the DINO family because DINOv3 produces brain-like activation patterns aligned with neuroanatomical structures. The four-backbone comparison demonstrates robustness to backbone choice. TracerAD has also been applied to a novel tracer currently undergoing regulatory review, though results cannot be disclosed at this time due to ongoing approval processes. The ADNI evaluation presented in this paper demonstrates the method’s feasibility on a publicly available dataset. We will explore 3D medical backbones and additional tracers in the journal extension.
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.
The paper presents an interesting approach for novel-tracer PET anomaly detection under few-shot settings. Additional justifications regarding methodological novelty and experimental evaluations could help further improve the quality of this paper.
