TL;DR
QUTCC introduces a nonlinear quantile calibration method for imaging inverse problems, producing tighter uncertainty bounds and better hallucination detection in medical and scientific imaging tasks.
Contribution
It proposes a novel nonlinear quantile calibration technique that improves uncertainty estimation in deep learning models for imaging inverse problems.
Findings
Tighter uncertainty bounds than prior methods.
Effective hallucination detection in medical imaging.
Maintains statistical coverage while reducing interval size.
Abstract
Deep learning models often hallucinate, producing realistic artifacts that are not truly present in the sample. This can have dire consequences for scientific and medical inverse problems, such as MRI and microscopy denoising, where accuracy is more important than perceptual quality. Uncertainty quantification techniques, such as conformal prediction, can pinpoint outliers and provide guarantees for image regression tasks, improving reliability. However, existing methods utilize a linear constant scaling factor to calibrate uncertainty bounds, resulting in larger, less informative bounds. We propose QUTCC, a quantile uncertainty training and calibration technique that enables nonlinear, non-uniform scaling of quantile predictions to enable tighter uncertainty estimates. Using a U-Net architecture with a quantile embedding, QUTCC enables the prediction of the full conditional…
Peer Reviews
Decision·Submitted to ICLR 2026
The paper presents a principled and rigorous approach to uncertainty estimation in imaging inverse problems. Its key contribution—the quantile-conditioned U-Net trained with a pinball loss and paired with nonlinear conformal calibration—represents a natural yet nontrivial extension of prior conformalized quantile regression methods such as Im2Im. The methodology is conceptually elegant and statistically grounded, providing a unified framework for pixel-wise uncertainty quantification, coverage g
While the paper is technically sound and well presented, several aspects could be improved to strengthen its contribution. First, the related work section omits an important body of literature on multi-modal prediction and multi-hypothesis uncertainty estimation using multi-head or mixture-based networks (e.g., Learning in an Uncertain World: Representing Ambiguity Through Multiple Hypotheses by Rupprecht et al., Hierarchical Uncertainty Exploration via Feedforward Posterior Trees by Nehme et al
Novelty and relevance: While the idea of adding conditioning mechanisms to enable a quantile regression network to learn the full quantile function is not novel (see [1, 2]), its application for uncertainty prediction in inverse imaging problems is novel, relevant, and timely. Extensive experiments: The authors conduct a thorough evaluation of their proposed method on five different datasets, selecting an appropriate baseline for comparison. References: [1] Ostrovski, Georg, Will Dabney, a
Major weaknesses: Insufficient improvements over baseline: My main criticism of this work is its minor (possibly zero?) improvements over the baseline Im2Im-UQ. The results shown in Table 1 and Figure 2 suggest nearly identical performance. Even in Figure 4, the selected outputs of the baseline and proposed methods are extremely similar. Consequently, it is difficult to justify that predicting the full quantile function versus 2 fixed quantile values (0.05 and 0.95) offers any real benefit. M
1. The manuscript is approachable and well written. The figures are clear. 2. The quantile embedding for image-to-image regression tasks is novel. 3. Estimating the conditional distribution and being able to mitigate quantile crossing offer practical utility.
1. The performance improvements compared to Im2Im-Deep (Table 1 and Table 2) are not entirely convincing and are marginal at best. 2. The crux of the method is the quantile embedding. However, the paper does not compare Im2Im-Deep on an equal playing field due to the different formulations (Im2Im-Deep uses symmetric, while QUTCC uses asymmetric). Specifically, it is unclear whether the performance gains are due to the asymmetric formulation, the quantile embedding, or both. It would make sense t
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
