acceptodds
Under review as a conference paper at ICLR 2027

To Control FDR, Constrain FNR: A Reversed Conformal Risk Training Framework

Abstract

In safety-critical image segmentation, controlling the False Discovery Rate (FDR) is essential. However, existing post-hoc Conformal Risk Control (CRC) for non-monotonic metrics like FDR yields overly conservative predictions, severely degrading model utility. While Conformal Risk Training (CRT) could theoretically mitigate this by embedding calibration into the training loop, it is fundamentally inapplicable to FDR. The inherent non-monotonicity of FDR violates the prerequisites for implicit differentiation, rendering the standard bi-level optimization intractable. To break this mathematical deadlock, we propose Reversed Conformal Risk Training (R-CRT). R-CRT employs a task reversal mechanism: it minimizes a differentiable surrogate of FDR while enforcing the strictly monotonic False Negative Rate (FNR) as a proxy constraint, thereby restoring the validity of implicit gradients. Extensive experiments on medical datasets demonstrate that R-CRT successfully trains highly effective models that significantly outperform baseline methods while maintaining strict, distribution-free FDR guarantees.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
conformal prediction, conformal risk control, conformal risk training, false discovery rate, medical image segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4JdZOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies medical image segmentation under an expected False Discovery Rate (FDR) constraint. It formulates the task as minimizing the False Negative Rate (FNR) subject to a user-specified FDR level. Since FDR is non-monotonic with respect to the prediction threshold, the authors argue that standard Conformal Risk Training (CRT) cannot be directly applied.

The paper proposes Reversed Conformal Risk Training (R-CRT). Instead of directly imposing the non-monotonic FDR constraint during training, R-CRT uses a monotonic FNR proxy constraint to determine per-sample thresholds, aggregates them with a soft quantile, propagates gradients through the threshold-selection step using implicit differentiation, and gradually tightens the proxy constraint through dynamic annealing. After training, the model is frozen and calibrated with Conservative Conformal Risk Control (CRC-C).

Experiments are conducted on several medical image segmentation datasets, with multiple backbones, target FDR levels, annealing strategies, threshold aggregation schemes, and controlled training baselines.

Critical strengths
  • The paper addresses a meaningful risk-constrained segmentation problem involving the trade-off between false discoveries and missed targets.

  • R-CRT provides a coherent algorithmic strategy for handling the non-monotonicity of FDR through a monotonic proxy, implicit differentiation, soft-quantile threshold aggregation, and dynamic annealing.

Critical weaknesses
  1. Insufficient positioning against prior work.
    The paper does not clearly establish its novelty relative to prior work on non-monotonic conformal risk control, particularly Angelopoulos (2026) and Aldirawi et al. (2026). These methods are mentioned but are not sufficiently discussed or directly compared. The experimental baselines also do not appear to cover all directly relevant methods for non-monotonic FDR control and conformal training.

  2. Incomplete theoretical justification of the reversed formulation.
    The paper does not establish a formal relationship between the proxy FNR constraint used during training and the target FDR constraint used after calibration, nor does it prove that the proxy optimization is equivalent to the original FDR-constrained objective. The main theoretical guarantee is adapted from prior CRC-C theory rather than being a new convergence, optimality, or generalization result for R-CRT. Moreover, the practical bootstrap estimate of the stability parameter is not a strict deterministic upper bound, so the implemented procedure does not directly instantiate the stated finite-sample guarantee.

  3. Incomplete isolation of the method’s components.
    Although the paper includes several ablations, they do not fully isolate the contributions of task reversal, implicit differentiation, dynamic annealing, and soft-quantile aggregation. The sensitivity to calibration-set size, quantile parameters, smoothing temperatures, and annealing choices is also not fully analyzed.

  4. Limited analysis of empirical behavior and theoretical assumptions.
    The experimental discussion is mostly descriptive. It provides limited analysis of failure cases, instability across datasets or data splits, the empirical behavior of the stability adjustment, and the relationship between the observed results and the assumptions underlying the theoretical guarantee.

Questions for authors

Please see Weaknesses.

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

This paper studies FDR control in image segmentation. Since FDR is not monotonic in the threshold, CRT cannot be applied directly. The authors propose R-CRT, which constrains the monotonic FNR during training while minimizing a smooth FDR loss. After training, CRC-C on a held-out set gives the final threshold. Experiments on several medical datasets show lower FNR than post-hoc CRC-C at similar FDR.

Critical strengths

The problem is practically meaningful, and FDR control in segmentation is less studied than FNR control, and the idea of using the monotonic FNR as a proxy is simple, cheap, and interesting. The improvement over the post-hoc baseline is fairly consistent across datasets and backbones.

Critical weaknesses
  • I do not fully understand why the reversal is valid. The original goal is to minimize FNR subject to FDR ≤ α. Training instead minimizes FDR subject to FNR ≤ α_FNR. The paper does not explain how α_FNR relates to the target α_FDR. For example, α_end is 0.2 in most settings but 0.9 for ACDC/BRISC at α = 0.1. An intuitive or theoretical explanation would help a lot.

  • The base models seem weak. For example, the base Dice on ACDC at α = 0.2 is only about 0.20, and HRNet on POLYP has FNR = 1.000. Could the authors report the base models’ performance at threshold 0.5 and test R-CRT on a stronger base model?

  • The contribution of the implicit gradient is unclear. The soft-FDR loss alone may already push the model toward higher precision. I suggest an ablation that stops the gradient through the threshold λ, i.e., no IFT term. This would show whether the bi-level / IFT part is really needed.

Questions for authors

See the weakness.

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

The paper aims to control the False Discovery Rate (FDR) in image segmentation while avoiding the near-empty masks often produced by post-hoc conformal calibration. Since Conformal Risk Training (CRT) requires a monotone constraint and FDR is non-monotone, the authors reverse the training objective: they constrain the monotone FNR while minimizing a differentiable FDR surrogate. The framework combines per-sample thresholds, soft-quantile aggregation, implicit differentiation, and an annealing schedule for the FNR target. After training, standard CRC-C is applied on a held-out calibration set to control the final FDR.

Critical strengths
  1. Practical and relevant problem: Post-hoc FDR control can lead to nearly empty masks, and the paper clearly illustrates this failure mode.
  2. Simple and efficient approach: The method is relatively easy to implement, with a short code listing and negligible additional computational cost (Tables 3–4).
  3. Broad experimental evaluation: The authors evaluate multiple datasets and backbones over 20 random splits, with useful ablations including extra training epochs and direct-loss baselines (Table 9), as well as single versus per-sample thresholds (Table 8).
Critical weaknesses

1.Limited novelty: The proposed method is essentially CRT with the objective and constraint swapped. The FDR guarantee comes entirely from the existing post-hoc CRC-C procedure, while the training stage itself provides no such guarantee. Moreover, there is no principled connection between the proxy level αFNR and the target αFDR. The annealing schedules (0.3→0.2, or 0.95→0.9 for certain datasets) appear to be manually selected for each dataset. 2.Missing baselines: Mossina & Friedrich (2025b), which addresses false-positive control in segmentation, is cited but not included in the comparisons. Learn-then-Test, a standard approach for handling non-monotone risks such as FDR, is not discussed. The claim that standard CRT fails when directly constraining FDR is supported only by the schematic in Figure 1, without a quantitative comparison. 3. Weak base models: The calibrated baseline FNR is extremely high in several settings: 1.000 for HRNet, around 0.90 for ACDC, and 0.94 for BRISC at α=0.1. This raises concerns about whether the base models are sufficiently trained or calibrated, potentially making the reported improvements easier to achieve. 4. Overstated guarantees: The stability parameter β is estimated using bootstrap rather than bounded deterministically, so the practical method does not provide a strict finite-sample guarantee. Although this limitation is acknowledged in Remark 1, the abstract and contributions still make strong claims about strictly satisfying FDR constraints. Several reported mean FDR values also exceed the target, including 0.205–0.207 for HRNet at α=0.2 and 0.102–0.103 at α=0.1. 5. Limited evidence for the annealing contribution: The four schedules (Fixed, Step, Exponential, and Cosine) perform similarly, with no consistently superior choice. In fact, Fixed performs best in some settings, such as ACDC at α=0.1 and POLYP at α=0.1 in Table 2. Thus, the claim that annealing provides an "optimal" FDR/FNR trade-off is not well supported. 6. Clarity issues: The train/calibration/test split is not sufficiently clear, particularly whether D_cal is fully disjoint from the R-CRT training data.

Questions for authors
  1. How are alphastart and alphaend selected without using test data? Is there a principled way to choose them?
  2. How does R-CRT compare with Mossina & Friedrich (2025b), Learn-then-Test, and stronger, fully converged base models trained with standard losses such as Dice or focal loss?
  3. Can the authors provide quantitative results for standard CRT with a direct FDR constraint, rather than relying only on Figure 1?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.