- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The authors present a novel method for estimating uncertainty for sparse-view GS-based CT reconstruction. Each Gaussian’s attenuation coefficient is modelled as an independent random variable, which leads to a closed-form uncertainty estimate due to the linear formulation of X-ray projection. Furthermore, the authors train a LightGBM error predictor on handcrafted voxel features, since the uncertainty maps correlate with error over the whole volume but insufficient within the anatomy. The predictor improves foreground error localization, driven mostly by image features and classical-reconstruction discrepancies rather than the uncertainty itself.
Critical strengths
- Relevant finding about foreground/ background
I find that the paper convincingly shows how whole-volume evaluation inflates uncertainty–error correlation (only 0.1 within body). This is an important caution for uncertainty evaluation in CT, and it holds across three uncertainty constructions and several foreground thresholds.
- Efficient uncertainty
The idea of making each Gaussian’s attenuation coefficient random exploits the linearity of X-ray projection. This is quite appealing, as it yields exact closed-form variance maps under the stated model, about 250× faster than Monte Carlo sampling, and the authors validate it against Monte Carlo estimates.
Critical weaknesses
- Feature design:
The 55 voxel-wise features (window sizes, scales, directional gradients) are not justified, and no feature-importance analysis is given. It remains unclear whether the predictor is mainly an edge detector. The authors need to elaborate on this.
- Limited and simulated evaluation:
The benchmark has only 15 scenes, several of them non-medical objects. The human experiments use simulated projections (which is ok), and there is no downstream task (e.g., lesion detection) or calibration analysis. Only ranking is assessed, not error magnitude.
- Clarity:
The paper would benefit from clearer terminology, such as distinguishing spatial Gaussian kernels from the attenuation distributions, consistent notation and a cleaner presentation of the experimental setups, particularly the diffusion comparison and the human-anatomy U-Net fusion.
Questions for authors
-
What are the main features contributing to the final estimation?
-
Could the authors evaluate the predicted error maps on a clinically motivated downstream task, for example by testing whether high-scoring regions coincide with missed or hallucinated lesions? Could they also report calibration of the predicted error magnitudes, rather than only rank-based metrics?
6 comments
Response to question 2
In terms of error magnitudes, the proposed learned error predictor will eventually give a score, which needs to be mapped to an actual error value.
Here is how we map scores to values in HU:
-
For each voxel in the development set, prepare a training pair consisting of its predicted score and its true error.
The ten MSD development subjects are split by subject into two groups of five. The tree model and the U-Net are trained on the 15 benchmark scenes and five of the subjects and used to score the other five. The two groups then swap roles. Each subject's scores thus come from a model that did not see that subject during training.
The normalized ranks from the two branches are averaged within the same scan to obtain the score . Since reference CT images are available for the development set, the true absolute error at that voxel can be computed directly as . Sampling 50,000 foreground voxels per subject gives a total of 500,000 pairs of scores and errors.
-
Use these paired data to learn a lookup table with 20 bins.
Since intensity scales differ across scans, first compute a scale for each subject as the 99th percentile of absolute intensity in the foreground of that subject's reconstruction. Divide the score range from 0 to 1 into 20 equal-width bins and compute the mean over the voxels in each bin.
Fit these 20 means using isotonic regression weighted by voxel count. If the estimates for two adjacent bins violate the ordering that “higher scores correspond to larger errors,” pool the bins and adjust their estimates. The resulting is the mean relative error for that score bin, learned from the development data.
-
Look up the value for a new scan and convert it back to HU.
For a new voxel, the model provides the score , and the reconstructed image provides the scale for that scan. The predicted absolute error is
Here, converts the relative error back to the scan's normalized intensity scale, and 4000 then converts the intensity difference to HU because the experiment maps a CT window of width 4000 HU to . Both test cohorts use the same lookup table. The reference CT images for the test subjects are used for the final check of the errors.
True and predicted absolute reconstruction error on 16 held-out CT scans (HU)
We applied the same frozen score-to-HU mapping to two disjoint cohorts of eight subjects each. Each cohort comprised four MSD and four LUNA16 subjects with simulated 25-view acquisitions. True is the absolute difference in HU between the reconstructed and reference CT values at a voxel. Predicted is the absolute error in HU estimated by the mapping for that same voxel.
The first column groups voxels by their fusion score . We used 20 equal-width score intervals to fit the mapping and 10 wider intervals to report the results. For each subject and score interval, we averaged the true errors and the predicted errors over exactly the same sampled foreground voxels. Each table entry is the median of the eight subject means in the indicated cohort, rounded to one decimal place.
| Fusion score | Original 8, true (HU) | Original 8, predicted (HU) | Additional 8, true (HU) | Additional 8, predicted (HU) |
|---|---|---|---|---|
| [0.0, 0.1) | 31.3 | 27.5 | 28.1 | 24.5 |
| [0.1, 0.2) | 37.1 | 36.4 | 34.9 | 32.4 |
| [0.2, 0.3) | 41.4 | 43.5 | 39.8 | 38.8 |
| [0.3, 0.4) | 47.7 | 50.1 | 44.5 | 44.7 |
| [0.4, 0.5) | 57.3 | 57.9 | 49.5 | 51.7 |
| [0.5, 0.6) | 71.1 | 67.9 | 58.4 | 60.6 |
| [0.6, 0.7) | 85.2 | 83.6 | 72.4 | 74.6 |
| [0.7, 0.8) | 107.9 | 107.8 | 96.4 | 96.1 |
| [0.8, 0.9) | 144.6 | 148.4 | 131.2 | 132.1 |
| [0.9, 1.0] | 229.7 | 254.1 | 232.1 | 226.9 |
| All sampled foreground voxels | 84.4 | 82.7 | 77.3 | 74.0 |
For example, in the highest score interval, the original cohort had a true mean absolute error of 229.7 HU and a predicted mean absolute error of 254.1 HU (subject medians). The corresponding values in the additional cohort were 232.1 HU and 226.9 HU. The final row pools all sampled foreground voxels within each subject before taking the cohort median.
To quantify calibration across the score range, we computed the absolute difference between the true and predicted bin means within each subject, then averaged these differences weighted by the number of voxels in each bin. The median of this subject-level calibration gap was 10.7 HU in the original cohort and 10.6 HU in the additional cohort.
We hope this would help clarify the calibration of predicted error magnitudes you are concerned about.
Reply to suggestions on the downstream task
Thank you for your valuable suggestion! Testing whether high-scoring regions coincide with missed or hallucinated lesions is a research area we haven't considered. The results will be included in our next paper, as it is already proving difficult to explain findings of this study within the conference’s page limit.
Best Wishes
Response to question 1
Hi, thank you for your recognition and valuable suggestions!
For your first question, features contributing most to the final estimation are normalized reconstruction intensity and standard deviation of local intensity in a 3×3×3 window (intensity means the X-ray attenuation value of each voxel). Top five features and their respective contributions are listed below.
We averaged within-model split-gain shares across 45 LightGBM models (15 leave-one-scene-out folds, three learner seeds). Fold ranks average the three seeds.
Average split gain share measures how much split gain a feature contributes when training a tree model. For example, a value of 7.42% for normalized reconstruction intensity means that, across 45 trained LightGBM models, the gain contributed by splits based on reconstruction intensity accounted for an average of 7.42% of the total split gain for each model.
Times entering top five indicates how many times, across 15 leave-one-scene-out folds, that feature’s contribution ranking fell within the top five. In each fold, contributions are recalculated.