acceptodds
Under review as a conference paper at ICLR 2027

Geodesic Conformal Risk Control for Provably Safe Image Segmentation

Abstract

Applying Conformal Risk Control (CRC) to image segmentation via standard pixel-wise thresholding often yields severely fragmented predictions under strict instance-level risk constraints. To address this, we propose Geodesic Conformal Risk Control (GeoCRC), a framework that ensures spatially coherent predictions while rigorously bounding the connected-component miss risk. GeoCRC leverages Isotonic Regression calibration to identify reliable seed regions and define traversal costs. Through multi-source geodesic propagation, it constructs a topologically well-behaved, nested family of regions parameterized by a geodesic radius, which is then calibrated using an instance-level miss loss. Theoretically, GeoCRC inherits the distribution-free marginal risk guarantees of standard CRC. Extensive experiments demonstrate that GeoCRC strictly satisfies user-specified risk constraints while drastically reducing spatial fragmentation and false-positive noise compared to baseline methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
conformal risk control, image segmentation, uncertainty quantification, geodesic propagation, instance-level risk control

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WLNeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Geodesic Conformal Risk Control (GeoCRC), a framework for image segmentation that aims to control instance-level miss risk while improving the spatial coherence of prediction sets. Compared to the original Conformal Risk Control (CRC), which is based on pixelwise thresholding, GeoCRC applies isotonic regression to calibrate pixelwise probabilities, extracts high-confidence seed regions, and finally constructs nested prediction sets through multi-source geodesic propagation. Experiments on four segmentation datasets with multiple backbone architectures show improvements in boundary accuracy and prediction fragmentation compared to the original CRC baseline.

Critical strengths
  1. The paper studies an important limitation of pixel-level risk metrics in image segmentation, more specifically, their potential to undersegment small target instances. The use of a connected-component-level miss loss is well motivated for applications where missing individual objects may be consequential.
  2. The proposed method combines probability calibration, seed-based geodesic propagation, and CRC in a conceptual pipeline while retaining the original CRC guarantee.
  3. The experiments cover four datasets and four segmentation architectures. The results show substantial reductions in boundary error and false positive components compared to the raw pixel-threshold CRC.
Critical weaknesses
  1. The main experiments primarily compare GeoCRC against raw pixelwise thresholding with CRC. The proposed approach incorporates spatial structure through seed selection and geodesic propagation, whereas the baseline does not exploit spatial relationships. This makes it difficult to determine whether the observed improvements arise from the specific geodesic formulation or simply from introducing spatial regularization.
  2. The framework mainly combines established techniques, including isotonic regression, geodesic propagation, and standard CRC calibration. The theoretical guarantee follows directly from the nestedness of the constructed prediction sets and the existing CRC theorem. While the combination is reasonable, the paper provides limited theoretical or empirical insight into why the particular geodesic construction should be preferable to other spatially structured prediction families.
Questions for authors
  1. Could the authors include more competitive spatially aware baselines, such as morphological post-processing or spatial smoothing combined with CRC, to better demonstrate the effectiveness of GeoCRC?
  2. ould the authors compare GeoCRC with simpler distance-based region-growing methods under the same input information and calibration settings to isolate the contribution of geodesic propagation?
  3. The authors acknowledge several spatially aware conformal methods (Bereska et al., 2025; Mossina et al., 2025; Gaillochet et al., 2026). Why are these methods not included as baselines, and could they be adapted to the same instance-level risk objective for a fair comparison?
Reviewer_yxbfOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Geodesic Conformal Risk Control (GeoCRC) as an alternative to standard Conformal Risk Control (CRC). The proposed framework successfully ensures spatially coherent predictions while rigorously bounding the connected-component miss risk for image segmentation, which is highly valuable for safety-critical domains such as medical imaging. The core concept is elegant, clear, and well-motivated. Furthermore, the empirical evidence presented in the paper demonstrates that this approach substantially improves performance compared with existing baselines.

Critical strengths

(1) The motivation and domain-specific benefits of this work are highly convincing. Although semantic segmentation is a well-studied field, spatial fragmentation and the disconnection or omission of small target instances remain persistent and unresolved challenges. This paper addresses a highly practical and critical research question for both medical diagnostics and industrial defect detection.

(2) The manuscript is exceptionally well-written and easy to follow. The conceptual framing and logical explanations are clear and straightforward. The primary claim is robust, well-supported, and can be summarized concisely. Theorem 1 provides a mathematical proof of the method’s coverage guarantee. Algorithm 1 precisely describes how the monotonic probability mapping and region construction and CRC calibration phases are implemented.

(3) The experimental evaluation is comprehensive and rigorous. The authors benchmark their method across four distinct datasets—Polyp, BRISC, ACDC, and Fire—using four baseline segmentation models. Their proposed model consistently outperforms the baselines across all evaluated scenarios, demonstrating substantial and meaningful gains over existing state-of-the-art frameworks. In Table 1, on BRISC with DeepLabV3(α=0.2), GeoCRC reduces ASSD from 76.556 to 6.236, HD95 from 145.885 to 13.501, and false-positive components from 125.548 to 0.059, while the reported component miss rate remains close to the target at 0.195. These results provide convincing evidence that changing the prediction family can substantially improve spatial quality relative to raw-threshold CRC. The validation firmly demonstrates the effectiveness of GeoCRC, and the underlying technical novelty offers valuable insights for the broader field of uncertainty quantification.

Critical weaknesses

(1) Contextualization within the geodesic-based segmentation literature: While the integration with Conformal Risk Control is unique, geodesic distance-based propagation has a long history in traditional and interactive image segmentation. The authors should broaden their literature review to highlight the precise methodological distinctions between their framework and existing classical or learning-based geodesic segmentation approaches.

(2) Analysis of computational cost: Although GeoCRC delivers significant empirical and topological advantages over standard CRC, multi-source minimum-cost geodesic propagation—for example, computing the shortest-path distance field—naturally introduces greater computational complexity than simple pixel-wise thresholding. Given the importance of operational efficiency for real-time, high-throughput imaging pipelines, the authors should explicitly report the inference time and memory overhead of their method.

Questions for authors

(1) How does GeoCRC compare conceptually and empirically with existing classical or learning-based geodesic segmentation methods? Was adapting geodesic propagation to satisfy CRC’s strict monotonicity and nesting criteria straightforward? If not, what was the most significant technical challenge encountered during this integration?

(2) How well does GeoCRC generalize to non-medical domains or structurally complex targets, such as human hair segmentation or detailed game-scene layouts? Based on the framework’s properties, which types of visual datasets are likely to benefit most from GeoCRC, and under what structural or noise conditions is it most prone to failure?

(3) Could the authors provide quantitative details on GeoCRC’s computational overhead? Specifically, how does it affect inference latency and memory consumption compared with standard raw probability-threshold CRC?

Reviewer_BRD4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeoCRC to control component-level miss risk in image segmentation while reducing fragmented predictions and false-positive noise. It constructs nested regions by geodesic expansion from seeds obtained using isotonic-mapped probabilities, then calibrates the expansion radius with standard CRC. Experiments across four datasets and four segmentation architectures report improvements over raw CRC at similar miss rates.

Critical strengths

The paper has a clear overall structure, with motivation, illustrations, and experiments presented in an accessible sequence, despite the presentation issues below.

Critical weaknesses
  1. Novelty and positioning. The broad priority claim at lines 74-75 needs careful qualification, especially because Section 2.2 already discusses spatial conformal methods. The seeded geodesic construction in Section 4.2 appears incremental, and Section 4.3 explicitly applies an existing CRC theorem. Please clarify what is new in the component-level objective and spatial construction, and how their combination advances beyond prior work.

  2. Related work and segmentation scope. Please discuss the relationship and a comparison to Spectral Conformal Risk Control: Distribution-Free Tail Guarantees via Bayesian Quadrature [1], including the distinction between tail-risk control and the expected component-miss objective here. The formulation uses binary masks. Please clarify the label preprocessing, especially for ACDC, and explain how the method would handle multiple semantic classes. Multiclass experiments would broaden the evidence, but comparisons must account for the different risk objectives.

  3. Limited baselines. Tables 1-2 compare only against raw CRC. Although Table 3 includes seed variants, a comparison with a relevant prior spatial conformal method under the same component-miss loss and comparable total labeled-data budget would better establish the method's advantage.

[1] Esfeh, Mohammad Mahdi Kazemi, et al. "Spectral Conformal Risk Control: Distribution-Free Tail Guarantees via Bayesian Quadrature." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Questions for authors

See above for major questions.

Minor presentation comments

  • Enlarge Figure 1 and its internal labels, and consider adding a separate ground-truth panel.
  • Tighten short line fragments at lines 79, 101, 147, and 171.
  • Define NN when the binary mask is introduced.
  • Correct duplicated equation-reference prefixes at line 161 and lines 236-237, and check the remaining references.
  • Consider placing table captions above Tables 1-3.
2 comments

Thank you for your response. I still have significant concerns regarding the novelty of the proposed idea and whether its contributions are substantial enough for a serious ML publication. In particular, the improvements appear quite incremental, and the original manuscript contains considerable weaknesses in its presentation and writing. Therefore, I will maintain my original score.

Thank you very much for taking the time to read our response and provide a follow-up. We genuinely appreciate your candid feedback, especially regarding the presentation and writing weaknesses in our original submission. Your comments pushed us to thoroughly revise the paper (highlighted in red) to improve its clarity, structure, and experimental rigor.

While we fully respect your perspective, we would like to briefly clarify why we believe our work offers a meaningful contribution rather than an incremental tweak:

  • Addressing a critical gap in safety-critical segmentation: Standard CRC in segmentation relies on pixel-level risks (dominated by large objects) and produces severely fragmented masks under strict constraints. GeoCRC shifts the objective to instance-level miss risk—treating small, critical targets equally—and resolves the resulting fragmentation through probability-guided geodesic propagation.
  • Verified beyond simple spatial post-processing: To directly address concerns about incrementalism, our revision incorporates controlled comparisons against spatial baselines (MorphCRC and SmoothCRC) calibrated under the exact same instance-level loss, along with additional benchmarks and ablations (Tables 1–7). These results demonstrate consistent improvements in boundary quality and spatial coherence that go beyond simple smoothing or morphological dilation.

We understand that the original presentation may have obscured these key points, and we hope the substantially revised manuscript addresses your concerns more clearly. We kindly ask if you might take another look at the revised version and reconsider your assessment, though we sincerely thank you for your time and thoughtful service regardless of the final outcome.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.