List of Papers Browse by Subject Areas Author List
Abstract
Dynamic positron emission tomography (PET) with tracer [18F]FDG enables non-invasive quantification of glucose metabolism by means of kinetic analysis, often modelled by the two-tissue compartment model (TCKM). However, voxel-wise kinetic parameter estimation using conventional methods is computationally intensive and limited by spatial resolution. Deep neural networks (DNNs) offer an alternative but require large training datasets and significant computational resources. To address these limitations, we propose a physiological neural representation based on implicit neural representations (INRs) for personalized kinetic parameter estimation. INRs, which learn continuous functions, allow for efficient, high-resolution parametric imaging with reduced data requirements. Our method also integrates anatomical priors from a 3D CT foundation model to enhance robustness and precision in kinetic modelling. We evaluate our approach on an [18F]FDG dynamic PET/CT dataset and compare it to state-of-the-art DNNs. Results demonstrate superior spatial resolution, lower mean-squared error, and improved anatomical consistency, particularly in tumour and highly vascularized regions. Our findings highlight the potential of INRs for personalized, data-efficient tracer kinetic modelling, enabling applications in tumour characterization, segmentation, and prognostic assessment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2825_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/tkartikay/PhysNRPET
Link to the Dataset(s)
N/A
BibTex
@InProceedings{TehKar_Physiological_MICCAI2025,
author = { Tehlan, Kartikay and Wendler, Thomas},
title = { { Physiological neural representation for personalised tracer kinetic parameter estimation from dynamic PET } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {497 -- 507}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors proposed to use implicit neural representations as personalized kinetic parameter estimation in dynamic PET, with anatomical CT priors. The proposed method demonstrated its data efficiency and reduced fitting error compared with the self-supervised DNN method.
- 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.
INR learns continuous functions, allow for efficient, high-resolution parametric imaging with reduced data requirements.
The anatomical priors from a 3D CT foundation model improved anatomical consistency and precision in kinetic modeling.
- 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.
It’s unclear that how many frames are with 2s, 30s, and 5min durations, and how many patients are split as training and evaluation.
The evaluation only contained configurations using 2D, 3D, high resolution, low resolution, and the incorporation with CT prior in comparison with the DNN-based method. However, a comparison with compartment model based parametric fitting and with parametric ground truth (simulation study) is needed.
The novelty might be limited since it seems a direct utilization of INR to kinetic modeling. The assumption of continuity might need to be further validated. The anatomical CT prior didn’t further improve quantative evaluations (Table 1).
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Please mention if the dynamic PET frames and the CT image are registered/motion corrected and the corresponding pipeline.
It’s unclear that FDB stands for the baseline method proposed by De Benetti et al.; please add a note in the text before first time using this term.
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The INR with anatomical prior idea seems interesting, but the improvement from anatomical prior seems not obvious and additional evaluations might be needed.
- 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 proposed using the implicit neural representations (INRs) combined with CT information for kinetic parameter estimation. The kinetic model was based on the conventional 2 tissue compartmental model that is usually used in dynamic FDG case. The proposed method was evaluated on 24 real dynamic scans with the long AFOV scanner. The proposed method demonstrated a lower MSE as compared to another self-supervised deep learning method.
- 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.
This paper introduced the a physiological neural representation for tracer kinetic modelling, and integrated the CT information to further improve the estimation. Overall, the method is novelty.
- 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 method description is not that clear. It is hard to imagine how the representation module and CT information help the kinetic fitting.
- It’s Okay to me without detail about kinetic output layer by the 2T model. But I am not sure other people will understand how use the k parameters to generate the time activity curve. This part make your method hard to follow.
- For FDG case, the most important parameter derived from the kinetic modeling should be the influx rate Ki. The phosphorylation rate k3 is interesting, but there is lack of clear application as compared to Ki. The important result regarding the Ki is missing in this paper.
- You mentioned multiple times that the proposed method outperformed the state-of-art, but only one comparison can not support this statement.
- I know there is no ground truth for k parameters in the real dynamic case, but only comparing MSE and SD is not sufficient, because those two metrics rely on your training strategy. There is a lack of quantitative comparison to demonstrate your k estimation is better or right.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
I have three additional comments for your consideration.
- Could you please make the dataset description clearer? What’s the image size, do you run the training and testing on the whole total-body image or just part?
- It’s better to include a traditional non-linear fitting result as your reference to show your estimation is reasonable. Now, only two DL methods can not show the reasonable range for each k parameter. 3 Do you have explanation about why your interest is kidney? Is there real application on that?
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
First of all, I would I am in the middle of weak accept and weak reject. However, considering the important k parameter result (FDG influx rate Ki) is missing, I made weak reject. I would be very looking forward to author’s rebuttal.
- 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 propose a patient-specific kinetic parameter estimation approach that does not rely on large datasets, using implicit neural representations for parametric PET imaging. This is particularly valuable in clinical scenarios where large-scale training data is unavailable. The proposed method demonstrates clear advantages over non-linear least squares, especially for voxel-wise parameter estimation, which is often noisy and unstable.
- 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.
Compared to conventional data-driven deep learning models, the proposed approach is data-efficient, avoiding the need for extensive pre-training. Furthermore, the integration of tissue kinetic modeling within the framework represents a sound contribution, enabling continuous and smooth function learning.
- 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 methodology presented falls under the broader umbrella of Physics-Informed Neural Networks. While the approach is implemented within the INR framework, there is existing literature demonstrating effective use of patient-specific PINNs for kinetic modeling. Notably, studies by Kong et al. and others have employed PINNs incorporating anatomical priors and direct modeling from raw signal domain data (e.g., sinograms), achieving reliable kinetic parameter estimation. In contrast, the current work uses an indirect parameter estimation approach. This represents a limitation in terms of methodological novelty, as it mainly showcases an application of INR.
Another key limitation is the lack of extensive validation reported. The results are demonstrated for only a single patient, which raises concerns about generalizability and robustness. Presenting kinetic parameter distributions (e.g., box plots) across a larger cohort (e.g., all 24 patients) would greatly strengthen the manuscript. While the use of CT-based foundation model features as informative priors is a promising idea, it would be far more impactful if integrated into a direct estimation pipeline.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
To justify the application impact reporting the kinetic parameter distribution of all 24 patients will be useful.
- 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.
PACS-based deployment is an interesting point made in the rebuttal, making clinical deployment more practical. Authors agree to provide a full 24-patient cohort, camera-ready, including box plots of every kinetic parameter for important organs.
Author Feedback
We thank the reviewers (R1, R2, R3) for their constructive feedback. Our responses are organised by theme.
1 Related work and approach As R1 notes, patient-specific PINNs have estimated kinetic parameters directly from sinograms or list-mode data. We choose an indirect route because these raw data are seldom stored in PACS, making clinical deployment impractical. To our knowledge, implicit neural representations (INRs) have not yet been applied to this problem; they offer higher spatio-temporal resolution and strong data-efficiency. Earlier PINN studies train on patient cohorts, whereas our network is deliberately patient-specific (one model per patient) and acts as a regularised solver. A conventional train/validation/test split (R2) therefore does not apply. In the HU-INR model, the CT intensity channel is fed to the INR together with spatial coordinates and time; this additional input enforces anatomical boundaries and improves spatial regularity of the estimated parameters (R2, R3). A fully sinogram-based, CT-guided pipeline is promising future work (R1).
2 Data A complete dataset description will be added in the camera-ready version (R2, R3). Each dynamic scan contains 2×10 s, 30×2 s, 4×10 s, 8×30 s, 4×60 s, 5×120 s and 9×300 s frames. Voxel size is 1.65 mm (isotropic); the matrix is 400 × 400 × up to 650, depending on patient height. We train on whole-body 3D volumes or on arbitrary 2D slices. Patients with visible motion were excluded (R2). PET and CT are coregistered by scanner design (R2).
3 Evaluation To move beyond the single case in the submission, we repeated every experiment for two additional patients (the rebuttal window allowed only these). Training stayed below 2 h, inference below 30 s and GPU memory below 13 GB, depending on image size. Adding CT Hounsfield units changed MSE by less than 2 % (R2). Classical voxel-wise non-linear curve fitting, requested by R2 and R3, was run on the original and the additional patients. Although its MSE is lower, inference can take up to 8 h per slice. In all three cases, curve fitting failed to optimise k3, leaving it near the initial guess, whereas the INRs retrieved k3 values within published ranges. The INR also yields more plausible VB, reaching about 35 % in the aorta versus at most 15 % with curve fitting. In the camera-ready paper we will process the full 24-patient cohort and, as R1 suggested, include box plots of every kinetic parameter for liver, spleen, kidney, aorta and possibly more. A simulation study would be uninformative because a patient-specific INR would fit synthetic data almost perfectly (R2).
4 Methodological clarity If accepted, we will add a graphical abstract (or schematic) and the exact equation that maps K1, k2, k3, VB and the input function to the predicted time-activity curve (R3).
5 Parameters R3 asked for the Patlak influx rate. From the two-tissue compartment model we know that Ki is given by: Ki = K1·k3⁄(k2 + k3) (Watabe 2016; Sari 2022). Since K1, k2 and k3 are already estimated, we will report Ki ranges per organ in the final manuscript.
6 Application [^18F]FDG-PET/CT is established for infection and inflammation imaging and is increasingly used in renal disease. Dynamic renal FDG modelling is a promising tool for quantifying renal glucose metabolism. Our method could therefore improve the diagnosis of renal infections, motivating our kidney focus (R3).
7 Minor clarification “FDB” refers to the baseline method of F. De Benetti et al. (2023) and will be expanded on first use (R2).
We believe these additions fully address the reviewers’ concerns and strengthen the manuscript.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This is a very interesting work and I’d like to invite the authors make the following two general improvements before the manuscript can be accepted for submission:
- The innovative idea of borrowing power from foundation models with PACS-based deployment would be more understandable with a diagram showing the methodology framework with more information about the system design. Only texture information will reduce the readability of the work.
- Provide the contents still missing but promised to reviewers that will be added in the camera-ready version. If possible, please open source the code too.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
All reviewers acknowledge the key contribution of this paper: using implicit neural representations for parametric PET imaging. The author/authors addressed the major concerns in rebuttal.