acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Computation with Conformal Certification

Abstract

Scientific imaging data involving medical (e.g., 3D radiographic, microscopic images) and astronomical sensors pose different challenges from natural images. One key challenge is that these datasets are often orders of magnitude larger than natural images. Efficient training and inference on such data is challenging due to the increased hardware requirements, often needing bespoke solutions. Moreover, the relevant objects such as tumors and astronomical phenomena are sparsely distributed with respect to the background information. Standard efficiency techniques like quantization and pruning apply uniform scaling of all regions, potentially compromising the precision/fidelity of predictions. Thus, efficient methods that reduce computational requirements without losing resolution for relevant parts of the data and with uncertainty estimates on predictions are essential. We present Conformal Adaptation with risk guarantees for uncertainty-aware sparse computations. Our formalism is general, spanning dense pixel-level classification and regression tasks where hierarchical models that predict at low resolution, certify the signal using conformal risk control, and refine the certified predictions. We benchmark our method on multiple tasks such as 3D medical image segmentation, 2D galaxy flux prediction, and 1D calcium spike rate prediction, achieving - wall-clock speedup across the various tasks, including on commodity CPUs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deep Learning, Conformal Prediction, Computation Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_boK2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a novel computation allocation paradigm: instead of speeding up each individual computation, it reduces the number of locations requiring costly computation, with CRC providing statistical guarantees for where computations can be skipped.

Critical strengths

It enjoys clear theory: admissible aggregation operators paired with CRC guarantee that the expected missed signal quality is bounded by α. It is highly generalizable, with tailored aggregation operators for segmentation and regression, and experiments spanning multiple scientific domains. It delivers substantial acceleration: FLOPs can be reduced by over 80×, with notable wall-clock speedups on both CPUs and GPUs.

Critical weaknesses

The paper emphasizes certified selection risk guarantees and uncertainty-aware sparse computation, yet what it actually guarantees is the amount of ground-truth signal contained within the discarded regions. Furthermore, the method has limited novelty, and it lacks fair comparisons against existing adaptive inference approaches.

Questions for authors

Beyond the aforementioned weaknesses, I note that the paper is missing citations and comparisons with several closely related approaches:

  • Conformal Risk Training: End-to-End Optimization of Conformal Risk Control
  • Enhancing image-conditional coverage in segmentation: adaptive thresholding via differentiable miscoverage loss
  • Automatically adaptive conformal risk control
6 comments
AuthorsOpenReview ↗

Rebuttal (Summary)

We thank the reviewer for their thoughtful review and for recognizing the theory, generality, and speedups. We appreciate the opportunity to strengthen our work with clarifications and provide additional baselines. We address the concerns as follows:

  • Guarantee: We clarified that CRC controls the expected fraction of ground-truth signal excluded from refinement, rather than final prediction accuracy.
  • Contribution: We clarified how our framework combines conformalized spatial selection, admissible label aggregation operators, and sparse execution. On HECKTOR, training both the coarse model and refiner requires 15$\times$ fewer FLOPs than Full, with DSC decreasing from 0.736 to 0.723, shows the consequence of our method goes beyond just inference-time improvements.
  • Additional comparisons: We added adaptive-selection, ROI-crop, and validation-DSC-tuned threshold comparisons. The results also show that CRC applies to different selection families.
  • Related work: We thank the reviewer for bringing the additional works to our attention and will cite and discuss Conformal Risk Training, COAT, and AA-CRC, explaining their connections to our framework.

We provide the detailed responses and results below for the reviewer's convenience.

AuthorsOpenReview ↗

Detailed comment (1/3)

W1 — What the guarantee covers. We thank the reviewer for pointing out this ambiguity. The reviewer’s interpretation is correct: CRC bounds the expected fraction of ground-truth signal contained in regions excluded from refinement by α\alpha. Here, “signal” is defined through the cell aggregation operator AA, allowing flexibility in what is measured.

The formulation is general across admissible definitions of signal: any nonnegative aggregation satisfying nullity and occupancy-preservation can be used, with the same CRC framework controlling the expected fraction of the resulting total weight excluded from refinement. Our wording may have obscured this flexibility, and we will make it explicit. This is the selection risk referred to in the paper; it does not guarantee final prediction accuracy. We will clarify this definition in the abstract and introduction and revise the phrase “uncertainty-aware sparse computation” to describe the calibrated missed-signal guarantee more precisely.

W2 — Novelty.

We thank the reviewer for highlighting these closely related approaches. We agree that cascades [1] and using conformal methods to guide computational allocation are established ideas. We have expanded the related-work discussion to make these connections explicit. Conformal early exit [3, 4] calibrates when intermediate predictions can be accepted, while C3PO [2] optimizes query-level model cascades under probabilistic cost constraints. Spatial methods such as ROI cascades and PointRend already decide where to compute.

Our contribution addresses spatial allocation within dense prediction: a coarse prediction selects regions for expensive fine-resolution computation, and CRC controls the expected fraction of ground-truth signal excluded from refinement. Our admissible label aggregation operators provide a common bounded, monotone loss formulation for segmentation occupancy and regression mass. We combine conformalized candidate selection with sparse execution and evaluate the framework across scientific domains, modalities, and sparsity regimes. Thus, our contribution is the formulation and execution framework for certified spatial refinement, rather than the general idea of a conformal cascade.

The computational benefits also extend to training. On HECKTOR, training the full pipeline—including both the coarse model and refiner—requires 15$\times$ fewer FLOPs than training the Full baseline, with DSC decreasing from 0.736 to 0.723. Our additional ROI and top-k experiments further demonstrate that the selection-risk formulation applies beyond the original refinement configuration. We will revise the introduction to make these contributions explicit.

References

[1] Viola, Paul, and Michael Jones. "Rapid object detection using a boosted cascade of simple features." Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001. Vol. 1. IEEE, 2001.

[2] Valkanas et al. 3po: Optimized large language model cas- cades with probabilistic cost constraints for reasoning.” Advances in Neural Information Processing Systems 38 (2026): 84481-84522.

[3] Jazbec, Metod, et al. "Fast yet safe: Early-exiting with risk control." Advances in Neural Information Processing Systems 37 (2024): 129825-129854.

[4] Schuster, Tal, et al. "Confident adaptive language modeling." Advances in Neural Information Processing Systems 35 (2022): 17456-17472.

AuthorsOpenReview ↗

Rebuttal (2/3)

W3 — Additional comparisons.

We thank the reviewer for providing us with the opportunity to improve our work by providing additional baseline comparisons.

(i) Adaptive selection. We adapt the selection principles of DynamicViT [5] and PointRend [6]. The original implementations do not directly support our 3D volume segmentation pipelines, so we adapt these selection principles within a shared pipeline, holding the coarse model, refinement network, and halo fixed. Top-k budgets match CRC's mean selected fraction on calibration data, isolating the allocation strategy from changes in architecture or training.

  • Score top-k (DynamicViT-inspired): Score top-k retains a fixed fraction of locations per sample, ranked by the coarse foreground / signal score. We use this as a proxy for DynamicViT, which learns token importance scores.
  • Uncertainty top-k (PointRend-inspired): prioritize uncertain locations, using the smallest ∣p−0.5∣\lvert p - 0.5 \rvert for segmentation and largest q95−q05q_{95} - q_{05} for regression. This adapts PointRend's uncertainty-prioritized refinement principle across tasks.
  • CRC top-k: calibrate the keep fraction of the score-ranked top-k family directly against the missed-signal loss. Its selections are nested, so CRC applies. This arm targets risk rather than matching the original selector's budget.

Missed-signal risk on held-out data at α=0.1\alpha = 0.1 (lower is better; the CRC top-k column gives the cells it needs):

TaskCRC threshold: cells refinedCRC thresholdScore top-k (DynamicViT-inspired)Uncertainty top-k (PointRend-inspired)CRC top-k
MSD U-Net2.1%0.0730.1120.530.076 (2.5%)
MSD Swin-UNETR1.7%0.0720.1370.600.077 (2.1%)
MSD VISTA3D2.2%0.0710.1330.680.077 (2.1%)
HECKTOR0.52%0.0860.0610.0620.086 (0.31%)
Galaxy2.2%0.0400.0380.780.094 (1.45%)
CASCADE8.5%0.0460.260.250.065 (22.7%)

MSD and HECKTOR use three seeds and 200 random half-splits of the original calibration scans, with calibration and evaluation separated within each split. These overlapping resplits assess sensitivity to calibration data; they are not 200 independent test datasets. Galaxy and CASCADE use held-out test sets. Budget matching is performed on calibration data.

Uncertainty-prioritized selection exceeds the target risk in five of six settings (0.25–0.78), except for HECKTOR. Uncertainty and signal coverage are different objectives. Score top-k improves on threshold CRC in HECKTOR and galaxy, but exceeds α\alpha on MSD and CASCADE.

Our method combined with top-k achieves mean held-out risk below α\alpha in every setting. On HECKTOR it uses 40% fewer cells than threshold CRC. The guarantee remains in expectation under exchangeability. Full results are in App. E.

Q1 — Related work. We thank the reviewer for highlighting these works. We will cite and discuss all three in Related Work and Sec. 8:

  • Conformal Risk Training (Yeh et al., NeurIPS 2025) differentiates through CRC during training and extends control to Optimized Certainty-Equivalent risks, including tail risks and downstream decisions. Our omitted-signal loss offers an objective for reducing selected volume at a fixed risk target.
  • COAT (Luo et al., ICLR 2026) learns image-specific thresholds through a differentiable miscoverage objective to improve image-conditional coverage. It naturally fits with our coarse model to improve input-adaptive spatial selection.
  • AA-CRC (Blot et al., AISTATS 2025) develops approximate conditional risk control with data-adapted conditioning functions and can improve the selection threshold from the coarse model.

These methods complement our use of calibrated selection to determine where an expensive refiner executes. Integrating them requires preserving their training and calibration assumptions, and provide avenues for significant and exciting future work.

References

[5] Rao, Yongming, et al. "Dynamicvit: Efficient vision transformers with dynamic token sparsification." Advances in Neural Information Processing Systems 34 (2021): 13937-13949.

[6] Kirillov, Alexander, et al. "Pointrend: Image segmentation as rendering." 2020 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2020.

Reviewer_ew6EOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a cascade method running in coarse and fine grain resolution. It uses conformal prediction guerantees to properly select the threshold for the coarse grained selection to control the missed predictions

Critical strengths
  • The topic it handles, fast inference when looking for a "needle in a haystack"
  • Good empirical results, large speedups with good relative performcen.
Critical weaknesses
  • Main issue is novelty, and important missing citations. Using a classifier cascade for speedups is a classical ML approach, e.g. Viola and Jones 2001. It is also not the first to combine a cascade with conformal predictions e.g. Valkanas et al 2025. The authors need to relate to these works and explain how exactly their method compares to other works
  • Missing important comparisons to other methods to speedup methods , and the ablation of using a fixed threshold for the coarse step. While there are some results in C.3 with fixed λ\lambda, it can't be compared directly. Find the optimal λ\lambda w.r.t DSC on the validation set and show how it compares to the adaptive conformal score.
  • Unreasonable speedup results. Unless I missed something you use a x4 grid so the best case scenario (as you use the same model) should be a 64 times speedup (just for running the coarse model), but you have x82 speedup in Tab. 1.
  • The writing should be improved. The method itself is quiet simple (which is a good thing) but section 3 explains it in a much more complicated way then it should be.
Questions for authors
  • What exactly is the novel contribution of this work considering all prior work on model cascades and its combination with conformal predictions?
6 comments
AuthorsOpenReview ↗

Rebuttal (Summary)

Reviewer 2 - working response

Summary Rebuttal

We thank the reviewer for their thoughtful review and actionable suggestions. We greatly appreciate the opportunity to improve our work. We have updated the manuscript to address the concerns as follows:

  • Contribution: We expanded the related-work discussion to clarify what our cascade certifies and at which granularity compute is allocated.
  • Additional comparisons: We added the requested validation-DSC-tuned threshold baseline, alongside ROI-crop, adaptive-selection, and reduced-precision comparisons. CRC is within 0.002 of the DSC-tuned rule in mean test DSC, with greater FLOPs savings.
  • Speedup accounting: We apologize for the confusion, and clarified how sliding-window inference explains the 82$\times$ speedup and added results without overlap.
  • Writing: We revised Sec. 3 to explain the method in three plain-language steps before introducing the general formalism.

We provide the detailed responses and results below for the reviewer's convenience. We again thank the reviewer for noting the emprical strength of our work, and providing us with the opportunities to improve and clarify our contribution.

AuthorsOpenReview ↗

Detailed comments (1/3)

W1/Q1 — Novelty.

We agree that coarse-to-fine cascades are a classical idea, and we should have cited Viola & Jones [1] and Valkanas et al. [2]. Both are now cited and discussed in Related Work, together with coarse-to-fine medical segmentation (e.g., the nnU-Net cascade). We do not claim the cascade as our contribution. Our contribution is what the cascade certifies and at which granularity compute is allocated.

Unit of adaptationHow the threshold is setGuarantee
Viola–Jones cascadeswhole window / sample (reject or pass)per-stage detection rate tuned empirically on validation datanone (empirical)
ROI cascades, PointRendspatial regions / uncertain pointsempirical or heuristic selectionno finite-sample bound on omitted ground-truth signal
C3PO (Valkanas et al.) [2]per query: exit or escalateconformal, label-freeprobabilistic cost control and generalization-error bounds
Conformal early exit (Schuster et al., Jazbec et al.) [3, 4]per sample: exit or escalateconformalagreement with the full model
Conformal Adaptationspatial cells inside one dense outputCRC on a label-based lossexpected fraction of true signal in the regions left unrefined

C3PO optimizes query-level LLM cascades using unlabeled model outputs, with probabilistic cost control and generalization-error guarantees. Spatial methods such as ROI cascades and PointRend already decide where to compute. Our contribution combines spatial allocation with a finite-sample bound on the fraction of ground-truth signal excluded from refinement. We also show training-cost reductions of up to 15$\times$ in our experiments.

Viola–Jones tune each stage to achieve a target detection rate on validation data. In our setting, CRC \textbf{automatically} selects a threshold with a finite-sample, distribution-free guarantee on the expected fraction of ground-truth signal excluded from refinement. As seen in Appendix C.3, a common uncalibrated cutoff does not transfer between tasks, where threshold λ=0.5\lambda = 0.5 misses 11% of the tumor cells on MSD but 86% on HECKTOR.

Our contribution is a task-general formulation of certified spatial compute allocation, a sparse execution scheme, and its evaluation across four scientific tasks. The new ROI-cascade experiment further shows that calibration can be applied to a conventional spatial cascade, beyond our original local-refinement implementation. We have revised the introduction to make this contribution explicit.

W3 — Why 82$\times$ exceeds 64$\times$.

We thank the reviewer for pointing out that the sliding-window setting was not clear. A 4$\times$ reduction along each spatial dimension reduces the voxel count by 64$\times$. The reported FLOPs, however, scale with the number of sliding-window evaluations. Both arms use the same 64$\times$64$\times$32 window with MONAI's default 25% overlap [5]. For HECKTOR, a 240$\times$240$\times$156 scan needs 175 overlapping tiles (2.6$\times$ redundancy), while the 60$\times$60$\times$39 coarse volume fits in 2 tiles. The coarse pass alone is therefore 87.5$\times$ cheaper in this setting, and refining the selected cells brings VISTA3D to 82$\times$.

As the ratio depends on the tiling, we now also report results with no overlap: 37.5$\times$ (VISTA3D), 25$\times$ (Swin-UNETR), and 1.2$\times$ (U-Net) FLOP reductions. We have added this breakdown below Table 1.

W4 — Writing.

We agree that the method is simple and apologize for obscuring this in Sec. 3. The revision opens Sec. 3 with three steps in plain words. For segmentation, these are: (1) run the model on a downsampled input; (2) select coarse cells whose foreground score exceeds a threshold calibrated by CRC; (3) refine only those cells and upsample the rest. The general formalism captures the shared structure across tasks without requiring a regular image grid or specific aggregation and refinement functions. The theorem and approach apply to both regression and classification problems.

References

[1] Viola, Paul, and Michael Jones. "Rapid object detection using a boosted cascade of simple features." Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001. Vol. 1. IEEE, 2001.

[2] Valkanas, Antonios, et al. "C3po: Optimized large language model cascades with probabilistic cost constraints for reasoning." Advances in Neural Information Processing Systems 38 (2026): 84481-84522.

[3] Jazbec, Metod, et al. "Fast yet safe: Early-exiting with risk control." Advances in Neural Information Processing Systems 37 (2024): 129825-129854.

[4] Schuster, Tal, et al. "Confident adaptive language modeling." Advances in Neural Information Processing Systems 35 (2022): 17456-17472.

[5] https://monai.readthedocs.io/en/0.9.0/inferers.html#sliding-window-inference

AuthorsOpenReview ↗

Detailed comments (2/3)

W2 — Additional baselines.

Following the reviewer's suggestions, we added validation-DSC-tuned thresholding, a conventional ROI-crop cascade, adaptive-selection comparisons, and a reduced-precision baseline.

(i) Fixed threshold tuned for DSC on the validation set. We thank the reviewer for the suggestion and apologize for App. C.3 not sufficiently answering this. On the calibration (validation) scans, we swept the same threshold grid CRC uses and picked the λ\lambda with the best Coarse + Refine DSC. We also picked the cheapest λ\lambda whose DSC is within ϵ\epsilon of that best. Each choice was then evaluated on the 96 held-out test scans (3 seeds):

ModelRule (chosen on validation)Test DSCCell miss rateFLOPs speedup
VISTA3DMax DSC0.7120.00018$\times$
Within ϵ=0.005\epsilon = 0.005 of max0.7070.108–0.13383$\times$
CRC, α=0.1\alpha = 0.10.7100.079–0.08982$\times$
Swin-UNETRMax DSC0.7020.027–0.04049$\times$
Within ϵ=0.005\epsilon = 0.005 of max0.6980.106–0.15456$\times$
CRC, α=0.1\alpha = 0.10.7010.090–0.09254$\times$
U-NetMax DSC0.6770.046–0.0592.0$\times$
Within ϵ=0.005\epsilon = 0.005 of max0.6730.150–0.2053.5$\times$
CRC, α=0.1\alpha = 0.10.6770.090–0.1052.7$\times$

Results are three-seed means; cell miss-rate ranges are across seeds. CRC is within 0.002 of the DSC-tuned rule in mean test DSC while reducing operation count by a further factor of approximately 4.5 for VISTA3D, 1.10 for Swin-UNETR, and 1.35 for U-Net. Maximizing DSC identifies an accuracy-oriented operating point, but does not explicitly target missed-signal risk or compute. Allowing ϵ=0.005\epsilon = 0.005 recovers greater speedups, with test miss rates reaching 0.205. All three rules choose from the same family of nested selections. CRC chooses its operating point from the interpretable target α\alpha; observed miss rates are 0.079–0.105 across nine runs. Its bound holds in expectation over calibration and test examples, rather than separately for every run. The full DSC–FLOPs curves are in App. E.

(ii) Adaptive selection. We adapt the selection principles of DynamicViT [6] and PointRend [7]. Score top-k retains a fixed fraction of locations per sample, ranked by the coarse foreground / signal score. We use this as a proxy for DynamicViT, which learns token importance scores. We also evaluate uncertainty top-k, inspired by PointRend. The original implementations do not directly support our 3D volume segmentation pipelines, so we adapt these selection principles within a shared pipeline, holding the coarse and refinement models fixed. We also combine top-k selection with CRC by calibrating the keep fraction.

Dataset / modelCRCScore top-k (DynamicViT-inspired)Uncertainty top-k (PointRend-inspired)
MSD U-Net0.073 (5%)0.112 (64%)0.525 (100%)
MSD Swin-UNETR0.072 (3%)0.137 (86%)0.599 (100%)
MSD VISTA3D0.071 (7%)0.133 (83%)0.676 (100%)
HECKTOR Swin-UNETR0.086 (33%)0.061 (7%)0.062 (8%)

(iii) Non-adaptive: reduced precision. Reduced precision changes the cost of network evaluation and can provide additional savings alongside spatial allocation. We tested BF16 autocast for the coarse model, the refiner and Full, with CRC re-calibrated under BF16 so that the certificate covers the deployed model.

ModelFull, BF16Coarse + Refine, FP32Coarse + Refine, BF16Mean DSC change (BF16)λ^\hat{\lambda} shift / miss rate (BF16)
VISTA3D1.4$\times$53$\times$63$\times$≤3×10−4\leq 3 \times 10^{-4}≤0.01\leq 0.01 / 0.084
Swin-UNETR1.0$\times$35$\times$33$\times$≤4×10−4\leq 4 \times 10^{-4}≤0.01\leq 0.01 / 0.092
U-Net0.8$\times$6.1$\times$5.9$\times$≤4×10−4\leq 4 \times 10^{-4}0 / 0.097
Reviewer_j8WHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Conformal Adaptation, a method for improving computational efficiency through a two-stage approach. A coarse model first predicts on downsampled, low-resolution input, and conformal risk control is used to select candidate cells for high-resolution refinement. The method is evaluated on brain tumor segmentation (MSD), head-and-neck tumor CT (HECKTOR), 1D calcium-imaging spike rates (CASCADE), , and 2D galaxy flux prediction, with 2.5-72x speedups.

Critical strengths
  1. The paper tackles an important and practical problem. Computational efficiency is a critical challenge in medical scanning. Spending compute only where the signal is present is a natural and useful idea.

  2. The method provides a distribution-free guarantee on the fraction of signal missed by the selection.

  3. The evaluation is broad. It covers MSD brain tumour, HECKTOR, 2D galaxy flux, and 1D CASCADE.

Critical weaknesses
  1. Efficiency is the paper’s main claimed benefit, but the paper does not define how FLOPs are counted. It does not say which operations are included, for example, downsampling, upsampling, sliding-window overlap, test-time augmentation, or the refiner.

  2. The paper does not describe the timing protocol. It is unclear whether the reported wall-clock time covers the entire inference procedure in Algorithm 3 or only the network forward passes, and whether the selected cells are refined individually or in batches.

  3. The paper reports speedups but not absolute run times. MSD reports only speedup ratios, while HECKTOR and CASCADE report only FLOPs. Without absolute run times, it is hard to tell whether the savings are on the order of minutes or milliseconds, or whether they are meaningful for the deployment settings discussed in the paper.

  4. The abstract claims wall-clock speedups “across the various tasks,” but wall-clock time is reported only for MSD and the galaxy task.

  5. On HECKTOR, the method misses more small lymph nodes than the full model, even though the guarantee is satisfied. The guarantee controls the total amount of missed signal, but may not protect small lesions. This is important because the paper identifies missing small lesions as a costly error. The paper attributes these misses to the coarse model, but does not analyze the issue further.

Questions for authors
  1. How are FLOPs counted for each task, and what tool or formula is used? Which parts of the pipeline are included in the FLOP count?

  2. Which steps of the inference are measured inside the wall-clock timing?

  3. What are the absolute runtimes of Full, and Coarse + Refine for each task, broken down by coarse pass, selection and refinement?

  4. Hardware is named only for MSD; which CPU and GPU were used for the galaxy timing, and why is wall-clock time not reported for HECKTOR and CASCADE?

  5. Are the missed lymph nodes mainly small lesions, and how does detection change with lesion size?

4 comments
AuthorsOpenReview ↗

Rebuttal (Summary)

Summary Rebuttal

We thank the reviewer for their thoughtful and insightful review, acknowledging the importance of our problem, our method for improving efficiency, and the breadth of our evaluation. Furthermore, we thank the reviewer for their constructive feedback and questions to improve our work. First and foremost, we sincerely apologize for omitting the compute-accounting and timing details.

We have updated our manuscript and provide detailed explanations to address the reviewer’s concerns. We summarize our detailed response here:

  • Timing details: We use FVCore [1] to measure FLOPs for the supported layers of each model tested in our experiments. We follow the convention of counting one fused multiply-add (FMA) or multiply-accumulate (MAC) as one FLOP. We also analytically verify the FLOP counts for our models. We now provide timing details and explicitly describe the operations included in these measurements. For all experiments, measurements include downsampling, coarse prediction, selection, gathering, refinement, and TTA. Only the initial I/O to move data from disk to VRAM and the final I/O to write results to disk are omitted.
  • Measured savings: We again apologize for omitting the absolute timings and have now included them for all tasks. We highlight results such as HECKTOR using eight CPU threads.
  • Small lesions: We thank the reviewer for suggesting additional analysis. We now provide a size-stratified analysis of lymph node detection, showing lower detection recall for smaller lesions. We have updated the limitations to clarify that controlling missed signal does not directly guarantee lesion-level recall. In particular, partial refinement of a lesion may be insufficient to satisfy the detection criterion. We believe improving coarse models for small lesions and exploring lesion-aware risks are potential solutions.
  • Hardware descriptions: We now include detailed hardware information for all experiments.

Further details of these amendments are provided in the revised manuscript and summarized in our point-by-point responses for the reviewer’s convenience.

[1] https://github.com/facebookresearch/fvcore

AuthorsOpenReview ↗

Detailed comment (1/2)

We reproduce the proposed updates in the text for the reviewers' convenience.

(W1/Q1) FLOP Counting: We thank the reviewer for the constructive feedback, and apologize for omitting the compute accounting details. We use the fvcore to measure the convolution, linear, attention, and normalization layers for the models. We use the fvcore convention where one multiply-add counts as a FLOP.

For the fully convolutional CASCADE and GalaxySim networks, we have exact control over the architecture and calculate an analytical FLOP count for these models as well, and the values agree with fvcore within 0.4%.

(W2/Q2) Timing protocol:

We apologize for the lack of detail for the timing protocol, and we updated Appendix D with the full timing details. The timer covers all of Alg. 3, from downsampling, coarse pass, upsampling, selection, refinement, including any test-time augmentation. Only the I/O placing data to and from the disk is omitted. All timings are repeated three times, with about 5% of the samples omitted for warm-up, and averaged over all test samples, unless otherwise explicitly noted. We also include the absolute runtimes.

(W3/Q3) Absolute runtimes: We sincerely apologize for accidentally omitting the absolute runtimes. Appendix D, has been added to include all absolute runtimes, and reproduced here for the reviewers convenience.

MSD (seconds per scan; GPU in ms):*

DeviceU-NetSwin-UNETRVISTA3D
CPU, 1 thread2.90 → 0.91 (3.2×)159 → 2.56 (62×)321 → 4.27 (75×)
CPU, 8 threads0.55 → 0.16 (3.5×)30.2 → 0.51 (59×)45.6 → 0.64 (72×)
CPU, 32 threads0.27 → 0.064 (4.3×)13.4 → 0.23 (58×)17.4 → 0.25 (70×)
L40S23.5 → 5.4 ms (4.4×)1090 → 17.7 ms (62×)1580 → 25.9 ms (61×)
A10039.1 → 7.3 ms (5.3×)1060 → 21.0 ms (51×)1220 → 24.6 ms (50×)
H100 NVL23.4 → 5.1 ms (4.6×)547 → 13.2 ms (41×)878 → 16.4 ms (53×)

MSD Coarse + Refine time by stage (CPU 1 thread / H100 NVL):

StageU-NetSwin-UNETRVISTA3D
Coarse pass0.30 s / 1.5 ms2.06 s / 10.2 ms3.82 s / 12.3 ms
Selection + refinement0.61 s / 3.7 ms0.50 s / 3.0 ms0.45 s / 4.1 ms

We group selection and refinement together for timing convenience.

Other tasks:

Task (unit)CPU 1 threadCPU 4 threadsCPU 16 threadsGPU
Galaxy (per field)12.6 → 0.32 s (40 ×\times)3.58 → 0.094 s (38$\times$)1.45 → 0.034 s (42$\times$)L40S: 85 → 2.9 ms (30$\times$)
CASCADE (per 256 windows)5.71 → 1.56 s (3.7$\times$)1.81 → 0.50 s (3.7$\times$)0.59 → 0.21 s (2.9$\times$)A40: 40 → 62 ms (0.65$\times$)
HECKTOR (per scan, 8$\times$ TTA)4.1 h → 3.5 min (71$\times$)8 threads: 32.5 min → 33 s (59$\times$)32 threads: 26 min → 19 s (83$\times$)H100-PCIe: 88.4 → 1.44 s (61$\times$); L40S: 155 → 2.1 s (74$\times$)

(W4/Q4) Reporting and hardware: Due to the heterogenous nature of our compute resource, we use a wide range of GPUs. For all CPU experiments, an Intel Xeon Gold 6348 (28C/56T) is used. For GPUs, we have:

TaskGPUs
MSDL40S, A100-80GB, H100-NVL
HECKTORL40S, H100-PCIe
GalaxySimL40S
CASCADEA40

All timings are performed on exclusive node allocations to minimize any shared cluster interference.

AuthorsOpenReview ↗

Detailed comments (2/2)

W5/Q5: Yes, the missed nodes have a median volume of 1.5 cm$^3$ compared to the 7.7 cm$^3$ for detected nodes. We have added a table in appendix C breaking down the recall vs volume (reproduced here).

GTVn volumenDetection recall
< 500 mm$^3$290.11
500–2000 mm$^3$630.33
2–10 cm$^3$710.61
> 10 cm$^3$630.80

here, n denotes the number of ground-truth lesions, and rates are averaged over three seeds. Interestingly, only 16% of the small lesions (<500 mm$^3$) have no selected cells for refinement. In other words, 84% of the lesions had at least one cell that was detected for refinement, but to total selection fell below the threshold of detection, as we require >50% coverage to count as detected.

As the reviewer correctly points out, the guarantee permits this because small lesions contribute fewer cells to the signal loss. The guarantee controls the expected per-scan fraction of occupied cells left unselected, not lesion detection recall. Our formalism does permit a lesion-aware risk loss instead that could give each lesion equal weight and control average lesion-weighted loss under CRC and is a strong extension. We highlight this limitation in our discussion section and again thank the reviewer for raising this important point.

AuthorsOpenReview ↗

Quick follow-up on rebuttal

Dear Reviewer,

Thank you again for your careful review and providing us with the opportunity to improve our work. We know this is a busy period, and we wanted to briefly follow up before the discussion period closes. Below is a short summary of the updates and clarifications in hopes to address your concerns:

  • FLOP accounting (W1/Q1): FLOPs are measured with fvcore for convolution, linear, attention, and normalization layers, counting one multiply-add as one FLOP. For the fully convolutional CASCADE and GalaxySim models, we also computed FLOPs analytically, and these agree with fvcore to within 0.4%.
  • Timing protocol (W2/Q2): The timer covers the whole of Algorithm 3: downsampling, the coarse pass, upsampling, selection, refinement, and test-time augmentation. Only disk I/O is excluded. Appendix D now gives the full protocol.
  • Absolute runtimes (W3/Q3, W4): We now report absolute wall-clock times for all four tasks across several CPU thread counts and GPUs, with a per-stage breakdown for MSD. For example, HECKTOR with 8× TTA goes from 32.5 min to 33 s on 8 CPU threads, and from 88.4 s to 1.44 s on an H100.
  • Hardware (Q4): We now list the hardware for every experiment: an Intel Xeon Gold 6348 for all CPU runs, and L40S, A100, H100 (different types) or A40 GPUs depending on the task. All timings were run on exclusive node allocations.
  • Small lesions (W5/Q5): We added a size-stratified recall analysis to Appendix C. Missed nodes are mostly small (median 1.5 cm³ vs. 7.7 cm³ for detected nodes). We also updated the limitations to state clearly that the guarantee controls missed signal, not lesion-level recall, and we describe lesion-weighted risks under CRC as a natural extension.

Since our rebuttal, we have also made two further improvements to the manuscript.

Additional baselines. We added four new comparisons:

  • Validation-tuned DSC threshold: CRC comes within 0.002 mean test DSC of a threshold chosen to maximize validation DSC. It also uses fewer FLOPs: about 4.5× fewer for VISTA3D, 1.10× for Swin-UNETR, and 1.35× for U-Net. Relaxing the DSC-tuned rule recovers similar speedups, but its miss rates are uncontrolled and reach 0.205. CRC's miss rates stay between 0.079 and 0.105 at the target α = 0.1.
  • Adaptive top-k selection (inspired by DynamicViT and PointRend): CRC outperforms both score-based and uncertainty-based top-k selection on all three MSD backbones. On HECKTOR, top-k selection is competitive. We report the full results.
  • Reduced precision (BF16): Reduced precision is complementary to our method. Combining BF16 with Coarse + Refine raises the VISTA3D speedup from 53× to 63×. Mean DSC changes by at most 4×10⁻⁴, and CRC re-calibrated under BF16 still meets the target (miss rates 0.084–0.097).
  • Conventional ROI-crop cascade: Calibrating the crop threshold with CRC, instead of fixing it at p > 0.5, gives higher DSC and larger speedups for all three backbones. For example, for U-Net the FLOPs speedup rises from 8.4× to 15.9× and DSC from 0.677 to 0.686. This shows that our framework also works on top of standard cascades.

Related work and novelty. We expanded the related work with more citations and now state explicitly how our approach differs from prior adaptive-computation and conformal methods.

We would be grateful to hear whether these revisions address your concerns, and we are happy to clarify anything else or answer further questions. We hope the changes resolve your main points, we would appreciate you considering them in your final assessment. Thank you again for your time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.