Anatomically-aware conformal prediction for medical image segmentation with random walks
Mélanie Gaillochet, Christian Desrosiers, Hervé Lombaert · 2026
Read Anatomically-aware conformal prediction for medical image segmentation with random walks by Mélanie Gaillochet, Christian Desrosiers, Hervé Lombaert (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-18997.txt.
Introduction
Mélanie Gaillochet, Christian Desrosiers, and Hervé Lombaert make error bars for medical image models respect anatomy. Doctors need to know how much to trust a model's outline of an organ or a tumor. Conformal prediction gives error bounds with strict odds backing, but plain use can yield broken, patchy sets.
Those patchy sets ignore the body's layout and are hard to use in clinic. The team adds random walks that follow anatomy, so the sets stay in one piece and match real structures. Trustworthy bounds that look like anatomy could help doctors act on model output with more care.
Catalog record
| CREATOR | Mélanie Gaillochet, Christian Desrosiers, Hervé Lombaert |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-18997
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Anatomically-aware conformal prediction for medical image segmentation with random walks Mélanie Gaillochet, Christian Desrosiers, Hervé Lombaert Abstract The reliable deployment of deep learning in medical imaging requires uncertainty quantification that provides rigorous error guarantees while remaining anatomically meaningful. Conformal prediction (CP) is a powerful distribution-free framework for constructing statistically valid prediction intervals. However, standard applications in segmentation often ignore anatomical context, resulting in fragmented, spatially incoherent, and over-segmented prediction sets that limit clinical utility. To bridge this gap, this paper proposes Random-Walk Conformal Prediction (RW-CP), a model-agnostic framework which can be added on top of any segmentation method. RW-CP enforces spatial coherence to generate anatomically valid sets. Our method constructs a k k -nearest neighbor graph from pre-trained vision foundation model features and applies a random walk to diffuse uncertainty. The random walk diffusion regularizes the non-conformity scores, making the prediction sets less sensitive to the conformal calibration parameter λ ^ \hat{\lambda} , ensuring more stable and continuous anatomical boundaries. RW-CP maintains rigorous marginal coverage while significantly improving segmentation quality. Evaluations on multi-modal public datasets show improvements of up to 35.4 % 35.4\% compared to standard CP baselines, given an allowable error rate of α = 0.1 \alpha=0.1 . 1 Introduction While deep neural networks can achieve impressive accuracy, they remain susceptible to failure when encountering distribution shifts, noise, or rare pathological cases. Quantifying and mitigating uncertainty is thus essential for building trust in automated segmentation pipelines. Uncertainty quantification (UQ) methods [ 13 , 19 , 16 , 22 , 34 , 1 , 18 ] capture aspects of predictive variability, but fail to provide valid prediction sets with theoretical guarantees and are often sensitive to both model design and data distribution. Conformal Prediction (CP) [ 4 ] has recently emerged as a powerful alternative that offers statistically valid prediction sets with finite-sample guarantees under minimal assumptions. Unlike heuristic uncertainty measures [ 13 , 16 , 22 ] , CP allows a user to specify a maximum allowable error rate, or miscoverage level, denoted by \alpha\in(0,1) . The framework then constructs a prediction set \hat{C}(X) that ensures marginal coverage, satisfying the condition P(Y\in\hat{C}(X))\geq 1-\alpha , where Y represents the true ground truth. In the context of medical segmentation, this guarantee ensures that the true anatomical structure is contained within the predicted region with a high probability. However, directly applying CP to pixel-wise outputs yields two critical limitations that hinder the utility of CP in clinical settings: a lack of spatial context and significant probability miscalibration. First, standard CP methods often ignore spatial context. Because the guarantee is marginal (averaged over pixels or images), it can result in noisy, fragmented regions and anatomically implausible boundaries. This is particularly problematic in medical imaging, where anatomical structures exhibit strong spatial continuity and clinical utility depends on an accurate, coherent representation of organ boundaries. Moreover, evaluations typically focus on statistical coverage guarantees. Practical evaluation metrics, such as the Dice score or Hausdorff distance, are rarely assessed. In practice, enforcing the coverage guarantee often comes at the cost of severe over-segmentation and excessive volume increase. This trade-off remains largely unquantified. Second, most CP methods derive their non-conformity score—a heuristic measure of prediction error, almost exclusively from the raw softmax probabilities of the model, which tend to be overly confident and mis-calibrated [ 16 ] . Foundation models [ 9 , 21 , 26 , 32 ] , trained on vast datasets, offer an opportunity to construct more informative and robust non-conformity scores for segmentation by leveraging their rich, high-dimensional feature embeddings. In this work, we propose a Random-Walk Conformal Prediction (RW-CP) framework for image segmentation to enforce spatial coherence of CP sets via the feature space of foundation models. Our method diffuses the predicted softmax probability map using a random-walk process guided by the feature space of foundation models. Probabilities are propagated through semantically similar neighborhoods, effectively regularizing and denoising uncertainty estimates. Following traditional conformal segmentation, we calibrate a global parameter \hat{\lambda} on these spatially-aware, diffused probabilities. The resulting segmentation region then retains the statistical guarantees of CP while better respecting anatomical boundaries. 2 Related work 2.1 Conformal prediction for medical image analysis The high-stakes nature of clinical applications requires statistically rigorous uncertainty quantification, making CP a natural fit for medical imaging. Early works in medical imaging primarily applied CP to whole-slide images to assess the confidence of segmented regions, such as in lung tissue analysis [ 35 ] . Other CP studies have focused on controlling image-level metrics [ 37 ] or improving calibration [ 8 , 10 ] . In particular, Conformal Performance Range Prediction [ 37 ] used CP to predict ranges of expected performance metrics, such as Dice score (DSC) or Intersection over Union (IoU), for image-level quality control. Kandinsky conformal prediction [ 8 ] clustered pixels based on the similarity of their non-conformity curve, in order to improve model calibration. 2.2 Conformal prediction for segmentation Few works have attempted to apply conformal prediction to segmentation tasks. Initial approaches focused on generalizing the probability thresholding method used in classification via conformal risk control [ 2 , 3 , 28 ] . Recently, [ 29 ] proposed to generate a margin around predicted mask, requiring only the predicted segmentation. To provide geometrically informed uncertainty quantification, Spatially-Adaptive Conformal Prediction [ 6 ] locally weighed the non-conformity scores according to the distance from key interfaces. Alternatively, Feature Conformal Prediction [ 33 ] constructed confidence sets within the deep feature space rather than the output layer to generate shorter confidence bands. However, all these methods face notable limitations in terms of input requirements or conformal set construction, including white-box access to model features [ 33 ] , specific image content [ 28 , 6 ] , or restriction to uniform margin expansion [ 29 ] . 2.3 Vision foundation models Vision Foundation Models (VFMs) based on self-supervised learning [ 9 , 30 , 32 ] , large-scale visual pre-training [ 21 , 26 ] , or general vision transformers [ 17 ] , are trained on massive, diverse datasets. VFMs are primarily developed on natural images [ 21 , 32 ] . While domain shifts typically necessitate adaptation modules to tailor these models to medical imagery [ 11 , 36 , 12 ] , recent evidence suggests that the features of VFMs are sufficiently robust for direct application [ 24 ] . Specifically, DINOv3 [ 32 ] has proven to be a highly effective off-the-shelf encoder for medical segmentation tasks. Our contribution This work addresses a critical gap in trustworthy medical image analysis: the lack of spatial coherence in uncertainty quantification. We introduce the Random-Walk Conformal Prediction (RW-CP) framework, the first split-CP approach for binary segmentation that leverages geometric diffusion to generate anatomically-informed valid prediction sets. Our framework is model-agnostic and enhances any base segmentation model by incorporating high-dimensional context from foundation models. We summarize our contributions as follows: • We introduce a novel scoring mechanism that combines foundation model features with random-walk diffusion to propagate uncertainty coherently across anatomical structures. • We provide a formal theoretical analysis to prove how regularizing the score-function gradient prevents set-size instability. • We extend the assessment of conformal sets beyond simple coverage guarantees to include segmentation accuracy .