List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Felix Meister, Chloé Audigier, Tiziano Passerini, Èric Lluch, Viorel Mihalef, Andreas Maier, Tommaso Mansi
Abstract
Radiofrequency ablation is a minimally-invasive therapy recommended for the treatment of primary and secondary liver cancer in early stages and when resection or transplantation is not feasible. To significantly reduce chances of local recurrences, accurate planning is required, which aims at finding a safe and feasible needle trajectory to an optimal electrode position achieving full coverage of the tumor as well as a safety margin. Computer-assisted algorithms, as an alternative to the time-consuming manual planning performed by the clinicians, commonly neglect the underlying physiology and rely on simplified, spherical or ellipsoidal ablation estimates. To drastically speed up biophysical simulations and enable patient-specific ablation planning, this work investigates the use of non-autoregressive operator learning. The proposed architecture, trained on 1,800 biophysics-based simulations, is able to match the heat distribution computed by a finite-difference solver with a root mean squared error of 0.51+-0.50°C and the estimated ablation zone with a mean dice score of 0.93+-0.05, while being over 100 times faster. When applied to single electrode automatic ablation planning on retrospective clinical data, our method achieves patient-specific results in less than 4 mins and closely matches the finite-difference-based planning, while being at least one order of magnitude faster. Run times are comparable to those of sphere-based planning while accounting for the perfusion of liver tissue and the heat sink effect of large vessels.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16449-1_17
SharedIt: https://rdcu.be/cVRUX
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a surrogate model for liver tumour radiofrequency ablation using a non-autoregressive operator learning
- Please list the main strengths of the paper; you should write about 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.
- Combining biophysical models and machine learning is a recent and developing area of research and the authors have made a good work in connection to planning of RFA
- Predictions were accurate and matching closely those using spherical models while taking into account patient-specific data
- Inference time similar to those of spherical models
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- Ground truth taken from simulation rather than real data.
- And to that point, it is unclear how the threshold set for generating the ground truth was selected.
- Validation dataset is small on only 10 patients
- Simulation parameters taken in general rather than from a distribution from possible patient-specific values
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
There are some details missing to be able to reproduce it fully but most of the architecture is described. However, datasets used for validation are not open.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
- It is unclear how much of the architecture is different from that of DeepONet
- Please clarify how inputs of the branch network are put together
- in relation to the trunk network, is it only the (elapsed) time used as input? if not, is that the set of queries starting from 0s? please clarify.
- Overall, it would be great to validate this approach on a phantom model with known material properties, geometry and heat sinks.
- Are there 14 (text) or 15 (figure) convolution networks?
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Very interesting paper backed up with novelty
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
Previous studies on automatic planning for Radiofrequency ablation commonly neglected the underlying physiology and used simplified spherical or ellipsoidal ablation estimates. The main contribution of this work was speeding-up biophysical simulations by using non-autoregressive operator learning, so that the proposed planning method could consider the biophysical effects of thermal ablation.
- Please list the main strengths of the paper; you should write about 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 work focused on an important challenge on the automatic planning of RFA, i.e., how to estimate ablation zone quickly during planning with considering biophysical effects.
(2) The authors proposed a noval method to enable fast, biophysics-based RFA planning. The authors considered the simulaiton as an “operator” and they used an operator learning approach. This is a new formulation for heat transfer simulaiton, which is interesting for MICCAI community, especially for computer assisted thermal ablation.
(3) The results of the proposed method were constent with the simulation using finite difference solver, which showed that the proposed method has the potential for fast estimation of theraml ablation.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
(1) The method part was not clear (some important information was not provided) and it is not easy for readers to follow this work.
(2) The results of the proposed method mainly depend on the reference solutions. However, there were two problems on the reference solutions: (1) formula 1 was not consistent with the formula in ref [2] and seems not right. (2) the simualtion resolution is 4mm, which seems too big for accurate simulaiton. The authors should clarify the two issues.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
It is not easy for readers to follow this work since the method part is not clear and no codes are provided.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
(1) The formula 1 seems not right as the advection term (a_vp_bc_bvgradT) and R(T_b-T) should be in seperated equations (see eq(2) and (3) in ref[2]). Please check it and make sure the reference solutions are right, otherwise the presented results may be wrong.
(2) The authors used 4mm grid for simulation, which seems not reasonable –4mm grid will make the geometry of liver and vessel distorted, thus the simulation was affected. For example, Fig 3 shows that the vessels are even not continuous, how did the authors simulate the blood velocity and make it right?
(3) The method part is not easy for readers to follow and the following information should be clarified:
a. it is not clear weather the input images are 2D or 3D. I guess it was 3D but the figure 1 showed the input images were 2D.
b. For the trucnk network, it is not clear weather the input was a time point or multiple time points.
c. What is the spherical source and what’s physical meaning of the radius of the spherical source?
minors:
typo->section 2.1 : for a(an) advenction-diffusion-reaction problem.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The idea of this paper is quite insteresting, and the results are good. However, the method part is not described clear and not easy for readers to follow.
- Number of papers in your stack
2
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
This work proposes the use of non-autoregressive operator learning approach for a very fast automatic RF liver ablation planning method.
- Please list the main strengths of the paper; you should write about 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 work is well justified and solid, with thorough evaluation. The authors address an important and interesting problem in the domain of treatment planning. The approach, which is much faster, has potential to significantly improve planning for RFA and treatment efficacy. .
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
The authors need to further clarify which portions of their work is their main contribution. For example in the introduction they state that the work was inspired by DeepONet. How much of this work is novel compared to prior work? This work was compared against one reference ([15]). But it is not clear from only that comparison where the current work stands in the literature.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Enough details are provided for repeatability in terms of parameters used. It is not clear if the code will be shared or the data would be available.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
The 4mm grid seems quite coarse. How would this affect results? A figure to illustrate the details in section 2.2. would be very helpful. It is difficult to visualize the approach through text only. Please further clarify what electrode position means in this application. Is it the depth of the needle? Coordinates of the tip of the needle? From Table 1 the superiority of the proposed method over the Sphere method is not clear. The latter does not account for the heat sink effect of blood vessels which I believe is the benefit of the proposed method over the Sphere method. Would there be a measure/metric to demonstrate that? Is AE sufficient as a comparison metric? Is it possible to understand in what situations the Sphere method is producing better AE % compared to the proposed method (in Table 1)
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Further clarification (see questions above) on novelty and results would greatly strengthen the paper.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Somewhat Confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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 manuscript presents a non-autoregressive operator in a deep learning framework to predict RF ablation volumes using biophysics parameters quickly and accurately.
The strengths of this work are that it is well justified, theoretically sound and the evaluation is thorough, although only on synthetic ablation volumes is used so the generalisability of this method to clinical data has yet to be demonstrated.
Weakness included how different this work is compared to DeepONet and other prior work within the literature. Also this work as it stands is not fully reproducible, as some details are missing and the code is not released.
Overall I suggest accept as the paper is considered to be technical interesting and innovative solving a complex and hard problem. While this work is still in early stages, using synthetic ablation volumes, it represents an innovative area of research.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
4
Author Feedback
Dear Reviewers,
We thank you for the positive response, the constructive feedbacks, and the interesting questions which we would like to address.
We would like to clarify Rev. #2’s concern about the correctness of Eq. 1. While ref [2] is indeed presenting a fully coupled biophysical model (Eq. 2 & 3), the actual model used in both the reference and our work is a two-compartment model (see Eq. 4 of ref. [2] and respective text). The reaction term is only applied inside the vessels and the advection term everywhere else. -> We revised the description and Eq. 1.
On the reviewers’ concern about the coarse grid resolution: indeed, a spacing of 4mm poses challenges in representing the anatomical structures. We agree with the reviewers and acknowledge that this may result in disconnected vessels and in turn deviations in the ablation estimates. As noted by Rev. #2, it is not possible to compute the blood flow in disconnected vessels. Therefore, we first compute the blood velocity at 1.3mm image resolution and then downsample it.
We further studied the effect of the grid resolution on the accuracy of estimating ablation zones with the finite-difference method (see our supplementary material). To this end, we first computed 25 ablation zones with randomized source location and radius on a single geometry at 4mm resolution. We then repeated these simulations at 1mm resolution by subdividing the voxels of the liver, vessel, and source mask, as well as the blood velocity field. Using the same upsampling scheme, we compared the 4mm ablation zone against the solution at 1mm and observed a dice score of 0.89 and an average Hausdorff Distance of 0.18mm, much smaller than the voxel spacing. Consequently, errors are only located at the ablation zone boundary. Thus, the 4mm solution is consistent with the 1mm solution as it captures the macroscopic behavior of the ablation zone subject to the advection and heat sink effect. Nevertheless, we agree that this is a particularly important issue to be addressed by future research. -> The data generation paragraph and the discussion were revised.
Addressing Rev. #3’s concern, we would like to clarify that our intention is not to show superiority of our method over sphere-based planning. We aim to show that our method can bring biophysical assumptions into the planning process at a fraction of the cost of a full computational model. We believe this has the potential to allow more precise and effective planning of ablation procedures especially in the presence of large blood vessels. Nonetheless, further research is necessary to demonstrate and quantify the superiority of biophysics-based models, e.g. with the help of large amount of intraoperative information, and to understand the impact of the blood vessels, the tumor shape and size, the planning objective, and modeling assumptions onto the planning results. -> The discussion was revised.
We would like to clarify our contribution w.r.t. DeepONet. The latter provides the generic framework for efficient operator learning, i.e. the separation of branch and trunk network as a consequence of the operator approximation theorem. We applied this framework to a complex, biophysical modeling problem, for which some of the inputs are naturally provided voxel-wise as they are based on medical images. To efficiently extract relevant features, we propose a specific convolutional branch network that enables fast and accurate emulation of RFA biophysics. -> A clarification was added to the final manuscript.
Short comments:
- Branch input: 3D volume with six features comprising the concatenation of the binary liver, vessel, and source masks as well as the blood velocity vector field
- Trunk input: A scalar, i.e. elapsed time (starting at 0s). Multiple queries are batched together for maximum efficiency
- Electrode position == coordinates of needle tip
- Spherical source == an approximation of the “umbrella-shaped” needle tip