- 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
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.
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.
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.
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).