acceptodds
Under review as a conference paper at ICLR 2027

Agentic Volumetric CT Reconstruction via Physics-Grounded Hypothesis Search

Abstract

Diffusion-based 3D CT reconstruction typically follows a single stochastic inference trajectory, limiting its ability to explore alternative volumetric solutions when measurements are incomplete. We introduce an agentic, physics-grounded framework that reformulates reconstruction as adaptive hypothesis search under a fixed test-time computation budget, without retraining the diffusion prior. A vision-language model (VLM) transfers cross-case reconstruction knowledge into a search prior that only prioritises which unexplored volumetric action should be tried first. Each action is realised by Hierarchical Physics-Compatible Evolution (HPCE), which combines shallow diffusion re-entry, full-volume data consistency, and physics-compatible innovation to generate structured 3D hypotheses whose diffusion-proposed structural changes remain constrained by the fit measurements. Realised hypotheses are evaluated using projection views held out from candidate formation, providing a case-specific, ground-truth-free physical state value. Physics-Grounded Monte Carlo Tree Search (PG-MCTS) uses this feedback to adaptively allocate the remaining computation budget across competing reconstruction trajectories. After search, we commit the selected action lineage rather than any search-stage volume, replay it using all acquired measurements, and accept each transition only when the full-measurement projection residual does not increase. Learning therefore prioritises where to search, held-out measurements determine what is worth pursuing, and full measurements govern final acceptance. Experiments on sparse-view and limited-angle CT, including cross-dataset evaluation, demonstrate consistent improvements over strong diffusion reconstruction baselines and matched-budget non-adaptive search, and a cross-task MRI study further shows gains over the diffusion baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Computed Tomography, Diffusion Models, Volumetric Reconstruction, Physics-Guided Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_X2yjOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a method that combines a diffusion prior, data splitting, a vision language model, and Monte Carlo Tree Search to effectively search for high quality CT reconstructions that fit the diffusion prior, the data, and cross-case prior knowledge through the VLM. The end result is a clever combination of methods, achieving higher accuracy metrics compared with comparison methods.

Critical strengths
  • clever combination of techniques from various subfields of CT, including data splitting for validation, diffusion priors for realistic imaging, VLMs for cross-case priors, and MCTS for efficient search strategies.
  • results seem to improve upon the state of the art.
Critical weaknesses
  • the method performs local exploration around the current state. It is unclear, however, whether this approach will still work when the probability distribution over reconstructions that fit the data is not local. For example, in medical CT, reconstructions that include a tumor and those that do not might both fit the data quite well, but the question becomes which case is more likely given the data. To answer this question, broad sampling of reconstruction space would be needed, while local exploration might get stuck into one state (i.e. with a tumor, or without). It would be good to discuss these settings in the paper, and ideally also include a specific experiment, e.g. taking one specific CT projection dataset, including an artificial low-contrast inclusion, and comparing results for both cases.
  • some information is only in the appendices, but could be better included in the main text, with more detail as well. For example, clear definitions of the B, C, and F actions are necessary in the main text. Another example is the computation time requirements, which is currently not very detailed in the appendix. Broader comparisons (with all comparison methods) on computation time and specific computational requirements should be included (e.g., is a H100 needed or is this also possible on more modest hardware?)
Questions for authors
  • how does your method handle rare cases and non-local probability distributions?
  • what are the specific computational requirements of your method?
  • what are the requirements of the capabilities of the diffusion prior and VLM?
3 comments

We thank the reviewer for their close reading of our paper, for highlighting the strengths of our work, and for their detailed and insightful comments. We have carefully considered all the points raised and address each one below.

W1/Q1: Local exploration, rare cases, and non-local reconstruction distributions

As described in Sec. 3.1, each HPCE step applies one of three actions (B, C, or F) to the current reconstruction to generate a new candidate volume. Different action sequences therefore explore different trajectories from a common initial reconstruction. PG-MCTS allocates its search budget between trajectories with stronger measurement support and branches that remain underexplored (Sec. 3.2). The range of anatomical alternatives explored depends on the candidates reached within this budget. At the end of the search, we select the candidate with the lowest held-out projection residual and replay its action sequence using all acquired measurements (Secs. 3.2–3.3).

In Fig. 3 of the revised manuscript, yellow dashed circles mark the same lesion location in the reference image and in each reconstruction of AAPM test case L506. The lesion visible in the reference image remains visible in the DDS+ISCS+Ours reconstruction.

Following the reviewer’s suggestion, we also compared reconstructions with and without a low-contrast target. In a relatively homogeneous region of the L506 reference volume, we inserted a 6-mm-diameter sphere by increasing the voxel values within it by 20 HU. We used the original volume as the target-absent control. We forward-projected both volumes at the same 30 view angles and reconstructed them with identical DDS+ISCS+Ours settings. We used the same target and surrounding-tissue ROIs throughout, defining local contrast as the mean intensity in the target ROI minus that in the surrounding-tissue ROI.

Local contrast (HU)Reference volumeFull-measurement initial reconstructionFinal reconstruction
Target absent1.01.51.1
Target present21.012.519.7

On L506, replay of the selected action sequence improved recovery of the 6-mm target’s local contrast. With the target present, local contrast increased from 12.5 HU in the initial reconstruction to 19.7 HU after replay, closer to the reference value of 21.0 HU. With the target absent, the final contrast at the same location was 1.1 HU, close to the reference value of 1.0 HU.

W2/Q2: Definitions of B, C, and F, reconstruction time, and hardware requirements

Thank you for this suggestion. Sec. 3.1 of the revised manuscript now defines the three HPCE actions (B, C, and F) and explains how each action updates the reconstruction.

We have added reconstruction times for all CT methods to Table 4 in Appendix C.2 and cited the table in Sec. 4.1. The table below reports the approximate time to reconstruct AAPM test case L506 (256 × 256 × 300 voxels, 30 views) on an NVIDIA H100. For DDS+Ours and DDS+ISCS+Ours, the times include both search and full-measurement replay.

MethodTime per volume (min)MethodTime per volume (min)
FDK0.1DDS+ISCS5
ADMM-TV2DiffusionBlend++13
DDNM5DDS+Ours10
DDS5DDS+ISCS+Ours10

To check whether the full method can run on a consumer GPU, we also reconstructed the same 30-view L506 volume on a single NVIDIA GeForce RTX 4090. With the same budget of 45 HPCE expansions, DDS+ISCS+Ours completed both search and full-measurement replay on the RTX 4090.

Q3: Required capabilities of the diffusion prior and VLM

The framework requires a diffusion prior that is suited to the target CT images and can resume denoising from a specified noise level. By varying the stochastic configuration through the HPCE actions, we can use the same prior to generate different candidate updates from the same reconstruction. The fit views constrain each update. Both our DDS and DDS+ISCS implementations use the same frozen, pretrained 2D diffusion prior (Secs. 3.1 and 4.1).

The VLM must be able to use the current reconstruction image and search context, including the search depth and action history, to prioritize which unexecuted HPCE action to try next. We adapt Qwen3-VL-8B-Instruct with LoRA using action preferences derived from the change in held-out projection residual after each action on the training cases. At test time, the VLM ranks the actions B, C, and F at the current node, and the highest-ranked unexecuted action is tried next when the node is expanded. Once a candidate has been generated, we evaluate it using its held-out projection residual (Sec. 3.2 and Appendix A.3).

Thank you for the response. These have strengthened the results in the paper.

Thank you for your thoughtful feedback. Your comments have helped us strengthen the paper.

Reviewer_BkMmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper reformulates diffusion-based volumetric CT reconstruction as measurement-guided trajectory search. Held-out projections guide computational allocation, and the selected action sequence is replayed using all measurements. The central idea is clear and meaningful, with encouraging reconstruction results. However, the evidence for the specific contribution of adaptive search is less convincing than the evidence for the overall framework.

Critical strengths
  1. Clear and relevant insight. Using held-out measurements to guide search, followed by full-measurement replay, addresses the tension between evaluating candidates and retaining scarce measurement information.

  2. Consistent empirical improvements. Gains appear across sparse-view and limited-angle CT, two related reconstruction backbones, and a separate dataset without retraining.

Critical weaknesses
  1. The contribution of adaptive search is not sufficiently isolated. The key ablations use a single patient, and the non-adaptive baseline is underspecified. It remains unclear whether the modest search advantage generalizes across patients and whether all reconstruction and replay procedures are matched.

  2. The value of the selected trajectory after replay remains uncertain. Replay changes both the starting reconstruction and the measurement constraints. Comparing replay against returning a search-stage volume also adds previously withheld measurements, so this comparison does not isolate the benefit of selecting a particular trajectory. Lower projection residuals alone do not resolve this concern.

Questions for authors
  1. Does the advantage over non-adaptive search persist across the existing LIDC test patients? Patient-level differences would substantially strengthen the central claim.

  2. Under identical full-measurement replay conditions, do search-selected trajectories outperform fixed or random trajectories of the same length? If an existing ablation already establishes this, please clarify which one.

1 comment

We thank the reviewer for their constructive suggestions and for highlighting the strengths of our work. We have carefully considered all the points raised and address each one below.

W1/Q1: Non-adaptive search baseline and patient-level results

As described in Sec. 4.1 and Appendix A.4, both adaptive and non-adaptive search use the same three HPCE actions (B, C, and F), a maximum search depth of 4, and a budget of 45 HPCE expansions. Both use the same fit/review split, initial reconstruction, and VLM rule for ranking actions that have yet to be tried at the current node. The two strategies differ in how they select the next node to expand. Non-adaptive search selects the shallowest eligible node, breaking ties according to a fixed ordering of action paths. This choice is independent of held-out projection feedback. Adaptive search uses UCT to select the next branch to explore based on held-out feedback from previously evaluated candidates. After search, each strategy selects the reconstruction with the lowest held-out projection residual from among the initial reconstruction and its own candidates, and retains the corresponding action sequence. Both strategies then replay their selected sequences from the same full-measurement initial reconstruction using the same physics acceptance rule.

We evaluated both strategies on the original 10 LIDC-IDRI test patients. For each patient, we computed the paired PSNR difference (adaptive minus non-adaptive) for each combination of imaging plane and random seed, then averaged these differences across the three planes and three seeds.

Test patientPSNR gain (dB)Test patientPSNR gain (dB)
1+0.546+0.62
2+0.577+0.50
3+0.668+0.65
4+0.689+0.61
5+0.5710+0.64

With the same budget of 45 HPCE expansions, adaptive search improved mean PSNR for all 10 patients. Patient-level gains ranged from 0.50 to 0.68 dB, with an average gain of 0.604 dB across patients.

W2/Q2: Search-selected vs. fixed or random trajectories under identical full-measurement replay conditions

In the 30-view AAPM experiment, our full method achieved a 0.61 dB gain in mean PSNR over the Return search-stage volume ablation, which returns the reconstruction generated using only the fit views (Table 2). To evaluate trajectory selection under identical full-measurement replay conditions, we compared search-selected, fixed, and random trajectories on the original 10 LIDC-IDRI test patients. For each search run, we matched the planned lengths of the fixed and random trajectories to the length K of the search-selected action sequence (K ≤ 4, Sec. 3.2). The fixed trajectory used the first K actions of the prespecified sequence (B,C,F,B). For the random trajectory, we sampled K actions uniformly from {B,C,F} with replacement.

Within each comparison, all three trajectories started from the same full-measurement initial reconstruction and used the same projections, HPCE settings, and physics acceptance rule. If a step increased the full-measurement projection residual, replay stopped and returned the most recently accepted reconstruction (Sec. 3.3). We evaluated the final reconstructions against ground truth.

TrajectoryMean PSNR (dB)Mean SSIM
Search-selected34.270.924
Prespecified fixed sequence33.020.880
Random sequence31.180.792

Under the same replay conditions, search-selected trajectories outperformed both controls. Compared with the fixed sequence, mean PSNR improved by 1.25 dB and mean SSIM by 0.044. Compared with random sequences, the corresponding gains were 3.09 dB and 0.132.

Reviewer_13AgOpenReview ↗
Rating
3
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose an agentic reconstruction pipeline for sparse view CT and limited angle view CT. In that framework they use a diffusion model to create a priors.

The research is potentially important as more and more low-dose and limited angle CT devices are hitting the market.

Critical strengths

I like about this paper is that the authors not only rely on PSNR, SSIM, MAE, RMSE,... for the evaluation of their reconstruction pipeline but also compare it to the ground truth directly using their |∆| metric. The exact unit of this metric is, however, unclear. I personally would expect that it is [HU] but given the super small values I assume it was determined in some AI friendly normailised range i.e. -1..1 or 0..1. This needs calrification. The measure is of course also heavily biased by the black background potentially dominating the image. It would be even more impressive/expressive if calculated excluding the background.

The agentic-based optimisation process looks interesting is thus not entirely new but probably new for agentic hypothesis searching.

Critical weaknesses

The runtime of 10min is quite long for such small volumes. The question is also how such an approach would scale to real world applications where scanners can capture images at different slice distances and such. Is one diffusion model enough to cover all these parameters?

For a physics-based approach it has to model many different effects such a attenuation, compton scattering, beam hardening and a lot more such effect. To what extent does a simplified approach help to mimic reality. This is unclear.

Questions for authors

The authors should clarify what the unit is in their evaluation metric |∆|. Although they shortly described it I can hardly believe that this in [HU]. I assume this number is given in some normalised gray-values. But this should be described.

1 comment

We thank the reviewer for their constructive suggestions and for highlighting the strengths of our work. We have carefully considered all the points raised and address each one below.

W1: Reconstruction time and slice spacing

On an NVIDIA H100, end-to-end reconstruction of the 256 × 256 × 300 AAPM test volume took approximately 10 minutes with DDS+ISCS+Ours and 13 minutes with DiffusionBlend++. When we increased the volume to 256 × 256 × 526 voxels using the same 30 projection angles, the corresponding times were 13 and 16 minutes. Our full method was therefore approximately 3 minutes faster than DiffusionBlend++ at both volume sizes.

At the input to our reconstruction pipeline, the 10 LIDC-IDRI test volumes had three distinct slice spacings ranging from 0.5 to 2.5 mm. During preprocessing, we resized each axial slice to 256 × 256 pixels and preserved the input slice spacing without through-plane resampling. We used the same frozen diffusion prior for all 10 volumes, with no spacing-specific retraining.

W2: Physical realism of the CT forward model

A common approach in conventional CT is to correct for effects such as scatter and beam hardening during projection preprocessing. At reconstruction time, the corrected, log-transformed measurements are treated as approximations to line integrals of the attenuation coefficient along the X-ray paths. Our framework follows this modeling convention.

To evaluate this approach on clinical measurements, we performed an additional experiment using the scanner-acquired projections released by Mayo/AAPM for case L506. These projections were exported after vendor preprocessing and the log transform, but before image reconstruction. All three methods used the same sparse-view projections and were evaluated against the full-view reconstruction of the same case.

On these measured projections, DDS+ISCS, DiffusionBlend++, and DDS+ISCS+Ours achieved PSNR values of 32.5, 33.1, and 34.7 dB, respectively, with corresponding SSIM values of 0.772, 0.875, and 0.918. Our method achieved higher PSNR and SSIM than both comparison methods on this case.

Q1: Units of ∣Δ∣|\Delta| and the effect of background

The ∣Δ∣|\Delta| values reported in our tables are dimensionless. As described in Appendix B.1, we clip CT intensities to [−1000,1600][-1000,1600] HU and linearly normalize them to [0,1][0,1]. In Eq. (19), xix_i denotes the ii-th slice on this normalized scale. We compute the mean absolute intensity difference between adjacent slices, SDiff, separately for the reconstructed volume and the ground-truth (GT) volume. We report ∣Δ∣|\Delta| as the absolute difference between these two SDiff values. A smaller value indicates that the reconstruction more closely reproduces the mean inter-slice intensity variation of the GT volume.

To assess the effect of background outside the body, we recomputed ∣Δ∣|\Delta| within the body region of AAPM test case L506. We derived a body mask from the GT volume that retained low-density regions inside the body and applied the same mask to all methods. After excluding the background, ∣Δ∣|\Delta| was 0.003068 for DDS+ISCS, 0.002801 for DiffusionBlend++, and 0.001770 for DDS+ISCS+Ours. As in the full-volume evaluation, DDS+ISCS+Ours had the lowest ∣Δ∣|\Delta| of the three methods and thus most closely matched the mean inter-slice intensity variation of the GT volume within the body.

Reviewer_NN9COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an agentic method using a self-defined Hierarchical Physics-Compatible Evolution to do 3D reconstruction of a CT scan. Instead of traditional single-trajectory diffusion methods, it uses different hypothetical reconstruction possibilities based on physics before committing to a single diffusion method. It does not rely on a single diffusion model as it uses a VLM and the Monte Carlo Tree Search to find the most promising possibility.

Critical strengths
  1. This paper identified and explained the problem of 3D CT reconstruction clearly, and the theory behind the method is well-defined with sound logic. In the Methods section, Figure 1 is relatively clear and showcases the described method well.
  2. The proposed agentic framework for 3D CT reconstruction for the closed-loop search is reasonable.
  3. The evaluation on multiple datasets shows consistent improvement on the CT reconstruction task.
Critical weaknesses
  1. This method is physics-based and calculates whether or not it fits into previous reconstructions, but would it become an inaccurate projection if there is an outlier in a given anatomy. It also depends on previous X-ray measurements, which may be incomplete.

  2. Using multiple hypotheses instead of previous diffusion-based methods does allow more options to improve accuracy, but it still does not guarantee that the options given are anatomically accurate. It also does not solve the problem of accepting inaccurate reconstructions that seem to be physically sound, but are actually inaccurate reconstructions

  3. In the Experiments section, Figure 2 does not clearly show or label an improvement, and it is hard to tell where it has visually improved even when looking at the diagram closely.

  4. The proposed method introduces the VLM to reorder the three predefined actions; after removing this design, the PSNR only drops by 0.19 dB. It is unknown whether introducing the VLM is feasible and whether the computational overhead is justified by the performance gains.

Questions for authors

How can the physics acceptance be guaranteed to be accurate instead of other diffusion methods as it still only uses previous reconstructions and given X-ray measurements?

5 comments

We thank the reviewer for their constructive feedback and for highlighting the strengths of our work. We have carefully considered all the points raised and address each one below.

W1: Anatomical outliers and incomplete X-ray measurements

For each case, we split the projection views into fit views, which are used to generate candidate reconstructions, and held-out review views, which are excluded from candidate generation. To evaluate a candidate volume xx, we compare its forward projection AreviewxA_{\mathrm{review}}x with the measurements yreviewy_{\mathrm{review}} for that case at the same held-out views (Sec. 3.2, Eq. 1). A smaller residual indicates that the candidate better explains measurements that were not used to generate it.

In the 30-view sparse-view CT experiment, for example, we use 24 views to generate candidates and hold out 6 for evaluation. After search, we use all 30 views, including the six held-out views, to construct the initial reconstruction and replay the selected action sequence. A replay step is accepted only if the projection residual over all 30 views does not increase (Sec. 3.3).

Regarding atypical anatomy, our AAPM test case L506 contains a metastatic lesion. Using the lesion location reported for L506 by Kim et al. [1], we added Fig. 3 to the revised manuscript and marked the corresponding region with a dashed yellow circle in both the ground-truth (GT) image and each reconstruction. The lesion visible in the GT image remains visible in our reconstruction.

We also evaluated 10 LIDC-IDRI cases with annotated pulmonary nodules, including two cases from the original test set and eight additional cases. All methods were evaluated under the same reconstruction settings. For each nodule, we used the same region of interest (ROI) in the GT image and in each reconstruction to compute local PSNR and SSIM.

MethodROI PSNR (dB) ↑ROI SSIM ↑
DDS+ISCS+Ours28.150.863
DDS+Ours27.620.820
DiffusionBlend++26.840.781
DDS+ISCS26.700.774
DDS25.750.756
DDNM25.430.729

DDS+ISCS+Ours achieved the highest PSNR and SSIM within the annotated nodule ROIs among the methods evaluated.

Reference

[1] Kim, K., El Fakhri, G., and Li, Q. (2017). Low-dose CT reconstruction using spatially encoded nonlocal penalty. Medical Physics, 44(10), e376–e390. DOI 10.1002/mp.12523.

W2: Anatomical accuracy of physically consistent reconstructions

During candidate generation, the fit views constrain the structural changes proposed by the diffusion prior. During search, we evaluate each candidate reconstruction using its projection residual on the held-out views (Sec. 3.2, Eq. 1). As stated in Sec. 3.2, “A smaller residual indicates that the reconstructed volume better explains measurements not used to form it.” During replay, an update is accepted only if the projection residual over all measurements does not increase (Sec. 3.3).

Because different reconstructions can fit the same measurements, we also compared each lesion region in the final reconstruction directly with the corresponding region in the GT image. In the newly added Fig. 3 of the revised manuscript, the metastatic lesion marked in the GT image for case L506 remains visible in our reconstruction. We also evaluated 10 LIDC-IDRI cases with annotated pulmonary nodules. Compared with DDS+ISCS, DDS+ISCS+Ours improved nodule ROI PSNR from 26.70 to 28.15 dB and ROI SSIM from 0.774 to 0.863. The table in our response to W1 reports the results for all methods.

W3: Visual improvements in Figure 2

Thank you for pointing out that the visual improvements were difficult to discern in the original Fig. 2. In Fig. 2 of the revised manuscript, a yellow box marks the region of interest in each axial, coronal, and sagittal image, and an inset in the upper-left corner of each panel shows a magnified view. Within each plane, the insets show the same region across all methods and the ground-truth image for direct comparison.

W4: VLM feasibility, computational overhead, and performance gains

At each search state, the VLM ranks the three predefined actions (B, C, and F) to determine which unexecuted action to try next (Sec. 3.2). In the 30-view AAPM experiment, adding the VLM improved both PSNR and inter-slice consistency at the same search budget (Table 2).

We repeated the ablation on the original 10 LIDC-IDRI test patients using three random seeds. Both variants used 45 HPCE expansions. The table reports mean PSNR across the 10 patients and three seeds, along with mean end-to-end reconstruction time per volume on NVIDIA H100 GPU.

MethodMean PSNR (dB) ↑HPCE expansionsTime per volume (s)
Full method (with VLM)34.2745600
Without VLM33.6845560

At the same search budget, VLM-guided action ordering improved mean PSNR by 0.59 dB at an additional cost of 40 seconds per volume.

Q1: What does physics acceptance guarantee compared with other diffusion methods?

Compared with the DDS and DDS+ISCS diffusion baselines used in our experiments, our framework adds two measurement-based checks. During search, each candidate is generated using only the fit views and evaluated on the held-out review views. A candidate that fits the views used to generate it can therefore receive a lower score if its predicted projections disagree with the held-out measurements (Sec. 3.2).

During replay, we forward-project both the current volume and the proposed candidate and compare their predicted projections with the same full set of measurements for that case. The current volume provides the baseline residual. We accept the candidate only if its residual is no greater than that of the current volume. Otherwise, replay stops and returns the current volume (Sec. 3.3). This guarantees that the final reconstruction’s residual over all measurements cannot exceed that of the initial reconstruction obtained using those same measurements.

Because different volumes can fit the same incomplete measurements, we assess anatomical accuracy separately against ground truth. In the 30-view AAPM ablation, retaining the acceptance rule yielded PSNR gains in all three anatomical planes and reduced the inter-slice consistency gap ∣Δ∣|\Delta| relative to replay without the rule (Table 2). For lesion regions, Fig. 3 of the revised manuscript and the LIDC-IDRI nodule ROI results in our response to W1 compare the full method’s final reconstructions directly with ground truth.

Thanks for the clarification and additional experiments. Based on your responses, there are two further concerns:

  1. The necessity of using a large VLM is still not sufficiently established. The additional experiments support that learning the action order is effective. However, the VLM is only in charge of the order of three pre-defined actions, lacking a comparison with a lightweight action-prediction model under the same search procedure, leaving it unclear whether the performance improvement indeed requires the large vision language model.

  2. The experiments provide limited evaluation of the method’s generalization. The main-paper ablation and AAPM results focus on a single case. Although the rebuttal includes experiments on 10 LIDC patients, the experiments using scanner-acquired projection data and the MRI experiments each cover only a single case. It remains a concern regarding robustness across different acquisition conditions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.