CARD: Calibration via Agreement in Reverse Diffusion for Out-of-Domain MRI Segmentation

Jiaheng Dai; Weidong Guo; Qingbiao Li; Jie Xu; Yi Guo; Yuanyuan Wang; Zeju Li

cs.CVAug 26, 2026

CARD: Calibration via Agreement in Reverse Diffusion for Out-of-Domain MRI Segmentation

Authors: Jiaheng Dai, Weidong Guo, Qingbiao Li, Jie Xu, Yi Guo, Yuanyuan Wang, Zeju Li

Organizations: College of Biomedical Engineering, Fudan University, Shanghai, China · Faculty of Information Science and Computing, University of Macau, Macau, China

Abstract

Probability calibration aligns model confidence with predictive accuracy, enabling clinicians to identify unreliable segmentation regions. This alignment breaks down under domain shift, where artifacts and unseen protocols produce confident errors. Existing post-hoc methods adapt the correction at test time, conditioning on predictive entropy, the logit pattern, or augmentation response, but each proxy is read from the terminal prediction, the very quantity that shift corrupts. This motivates reliability evidence beyond the terminal prediction, which categorical diffusion provides in two ways. First, a generative shape prior keeps a capacity-limited reference intact when appearance is corrupted, so its disagreement with the primary segmentor highlights primary-model errors. Second, every reverse step yields a class distribution, separating persistent disagreement from transient discrepancy. Aggregated over the trajectory, this disagreement correlates with Dice at 0.788, against 0.521 for a matched discriminative control. We therefore propose CARD (Calibration via Agreement in Reverse Diffusion), which maps the temporal aggregate of this disagreement to a temperature field applied per pixel across all classes, so that confidence changes while the segmentation does not. Across cardiac, prostate and brain MRI shifts, CARD lowers calibration error in 45 of 49 comparisons against the strongest baseline in each setting.

Figures & tables

Fig. 1: CARD derives reliability evidence from the prediction process. (a) Existing families take their evidence from the final prediction, or from nothing at all: static methods apply a frozen correction, output-conditioned methods read the terminal output, and image-conditioned methods add augmented views and a shape prior. CARD instead compares a primary and a capacity-limited reference across shared reverse states, the only evidence source in the table that is not terminal. (b) NLL on clean cardiac MRI and four artifact shifts. CARD stays lowest across every corruption, and the margin widens with severity.

Fig. 2: Disagreement tracks accuracy because only the primary fragments. (a) Largest connected component ratio. The primary loses connectivity as corruption worsens while the reference stays intact, so disagreement marks the primary’s failures, not the reference’s. (b) Slice-level agreement against Dice under motion corruption, with two matched discriminative U-Nets as a control. (c) Pearson correlation across four artifact types and four vendor groups.

Fig. 3: Overview of CARD. (a) Dual reverse diffusion. From input MRI $\bm{I}$ , the primary and reference denoisers follow paired DDIM-style reverse trajectories, initialized from the same categorical noise state $\bm{x}_{T}$ and evolving through intermediate states $\bm{x}_{t}^{\theta}$ and $\bm{x}_{t}^{\phi}$ to predicted clean states $\hat{\bm{x}}_{0}^{\theta}$ and $\hat{\bm{x}}_{0}^{\phi}$ , respectively. At each step $t$ , these states define the pixel-wise categorical distributions $\hat{\bm{p}}^{\theta}_{t}(\bm{v})$ and $\hat{\bm{p}}^{\phi}_{t}(\bm{v})$ over $K$ classes. Their pixel-wise disagreement is denoted $d_{t}(\bm{v})$ and computed by JS divergence. (b) Temporal aggregation. We aggregate ${d_{t}(\bm{v})}_{t\in\mathcal{T}}$ over the sampled reverse states to obtain a stable pixel-wise signal. (c) Pixel-wise temperature scaling. The aggregated disagreement defines a temperature map $\tau(\bm{v})$ , which calibrates the final primary logits by $\mathrm{Softmax}(\bm{z}(\bm{v})/\tau(\bm{v}))$ .

	(a) Cardiac artifact and cross-vendor shifts
	ACDC-C artifact						M&Ms vendor
Method	Clean (ID)	Bias	Mot.	Ghost	Spike	NLL $\downarrow$	A (Sie.)	B (Phi.)	C (GE)	D (Can.)	NLL $\downarrow$
No post-hoc calibration
UC	6.36/3.28/2.71	11.40/5.89/5.34	9.75/5.00/4.29	11.40/5.86/5.24	24.00/12.31/11.75	0.945	16.79/8.61/7.87	10.16/5.20/4.45	8.48/4.37/3.67	11.09/5.74/5.04	0.792
Alea. [ 27 ]	5.96/3.06/2.55	10.95/5.66/5.15	10.10/5.20/4.54	10.49/5.41/4.80	23.64/12.05/11.50	1.037	16.69/8.52/7.82	10.01/5.13/4.47	8.23/4.22/3.61	10.52/5.39/4.72	0.864
Static

Table 4

Fig. 4: Trajectory disagreement lands where the primary is wrong, and calibrating on it closes the confidence gap. (a) Ground truth, primary and reference predictions, and the temporally aggregated JS disagreement for a ghosting-corrupted cardiac case. The disagreement map coincides with the primary’s segmentation error in (d) , without access to labels. (b) Reliability diagrams for the five calibration methods; shaded bars give the confidence-accuracy gap. (c) Confidence histograms of the corresponding predictions. (d) Primary segmentation error against the ground truth. (e) Ghosting-corrupted input image with ground-truth contour.

Fig. 5: Calibration favours a small, undertrained reference: the best NLL in every task comes from the smallest model, well before convergence. Panels (a) to (c): cardiac, prostate, and brain. Rows are reference models with parameter counts, columns are training steps, and cells report $\mathrm{NLL}\times 100$ (lower is better, with the best cell per task outlined). The Base reference degrades steadily with further training, consistent with a high-capacity reference reproducing the primary’s failures and reducing prediction contrast. Colors are normalized within each panel.

Uncertainty Calibration

Uncertainty calibration aligns model confidence and predictive intervals with observed correctness and outcome frequencies.

See all 265 papers

Confidence-Calibrating Regularization for Robust Brain MRI Segmentation Under Domain Shift
Sep 27, 2025Behraj Khan, Tahir Qasim Syed, Syed Ahmad Chan BukhariImage Segmentation Promptable Image Segmentation
St. John’s University New York, USA · Institute of Business Administration Karachi Pakistan
Rethinking Post-Hoc Calibration in Semantic Segmentation
Jul 2, 2026Tristan Kirscher, Kim-Celine Kahl, Balint Kovacs +5Post-Hoc Calibration Model Calibration
ICube, Institut Strauss · ICube Laboratory, CNRS UMR 7357, University of Strasbourg, Strasbourg, France · CLCC Institut Strauss, Strasbourg, France +6
SNR-Adaptive Unified Diffusion for Multi-Task Medical Image Segmentation
Jul 3, 2026Jiahao Liu, Hang Wei, Shuai WuUnsupervised Domain Adaptation Domain Generalization
School of Telecommunications Engineering, Xidian University · School of Computer Science and Technology, Xidian University

Sep 27, 2025J/K move · Enter open · S save

Confidence-Calibrating Regularization for Robust Brain MRI Segmentation Under Domain Shift

Behraj Khan, Tahir Qasim Syed, Syed Ahmad Chan Bukhari

St. John’s University New York, USA · Institute of Business Administration Karachi Pakistan

Image Segmentation Promptable Image Segmentation Domain Generalization Medical Image Segmentation 3D Medical Image Segmentation Segmentation Uncertainty Estimation

Open paper AskPDF arXiv

The Segment Anything Model (SAM) exhibits strong zero-shot performance on natural images but suffers from domain shift and overconfidence when applied to medical volumes. We propose CalSAM, a lightweight adaptation framework that (i) reduces encoder sensitivity to domain shift via a Feature Fisher Information Penalty (FIP) computed on 3D feature maps and (ii) penalizes overconfident voxel-wise errors through a Confidence Misalignment Penalty (CMP). The combined loss,

\mathcal{L}_{\mathrm{CalSAM}}

fine-tunes only the mask decoder while keeping SAM's encoders frozen. On cross-center and scanner-shift evaluations, CalSAM substantially improves accuracy and calibration: e.g., on the BraTS scanner split (Siemens

\to

GE) CalSAM shows a

+7.4\%

relative improvement in

\mathrm{DSC}

(80.1% vs.\ 74.6%), a

-26.9\%

reduction in

\mathrm{HD95}

(4.6 mm vs.\ 6.3 mm), and a

-39.5\%

reduction in

\mathrm{ECE}

(5.2% vs.\ 8.6%). On ATLAS-C (motion corruptions), CalSAM achieves a

+5.3\%

relative improvement in

\mathrm{DSC}

(75.9%) and a

-32.6\%

reduction in

\mathrm{ECE}

(5.8%). Ablations show FIP and CMP contribute complementary gains (

p<0.01

), and the Fisher penalty incurs a modest

\sim

15% training-time overhead. CalSAM therefore delivers improved domain generalization and better-calibrated uncertainty estimates for brain MRI segmentation, while retaining the computational benefits of freezing SAM's encoder.

Abstract

Uncertainty Calibration

Explore similar work

Confidence-Calibrating Regularization for Robust Brain MRI Segmentation Under Domain Shift

Rethinking Post-Hoc Calibration in Semantic Segmentation

SNR-Adaptive Unified Diffusion for Multi-Task Medical Image Segmentation