Accurate MRI-free PET Quantification and Reliability Assessment via CT-to-MRI Synthesis and Conformal Prediction

Jungwoo Seo; Seung Yun Choi; Kun Ho Lee; Seun Jeon; Byoung Seok Ye; Jiook Cha

cs.CVOct 29, 2026

Accurate MRI-free PET Quantification and Reliability Assessment via CT-to-MRI Synthesis and Conformal Prediction

Authors: Jungwoo Seo, Seung Yun Choi, Kun Ho Lee, Seun Jeon, Byoung Seok Ye, Jiook Cha

Organizations: Seoul National University, Seoul, Republic of Korea · Yonsei University, Seoul, Republic of Korea

Abstract

Standardized uptake value ratio (SUVR) quantification of amyloid PET commonly relies on a subject-specific T1-weighted MRI, yet a usable MRI is not always available. We therefore combined CT-to-MRI synthesis with conformal prediction for MRI-free PET quantification with reliability assessment. A slice-consistent Brownian Bridge Diffusion Model was strengthened with multi-dataset training and data-source conditioning to enrich the anatomical prior learned from MRI, and with expanded multi-slice CT context to exploit the through-plane continuity of subject-specific anatomy in CT. Synthetic MRI was used for FreeSurfer-based anatomical parcellation and PET quantification, while deep-ensemble uncertainty was calibrated with conformal prediction to estimate a plausible range for the corresponding real MRI-based SUVR. On $^{11}$ C-PiB PET from OASIS-3, the proposed method achieved the highest synthesis and parcellation fidelity among the compared methods. Mean cortical SUVR from the proposed synthetic MRI agreed closely with the real MRI-based reference (MAE $0.016 \pm 0.015$ , Pearson $r = 0.9987$ ) and yielded the same amyloid-positivity classification as the real MRI in all 87 test subjects at the prespecified cut-off. Conformal prediction achieved empirical coverage close to the prespecified levels, and at 90% target coverage, 85 of 87 subjects had prediction intervals that supported an unambiguous amyloid-positivity decision. Zero-shot evaluation on an independent external cohort showed that the proposed CT-to-MRI model could support PET quantification without retraining, although conformal calibration required target-cohort recalibration to recover the intended coverage. These findings support CT-to-MRI-based MRI-free PET quantification as a promising approach for providing accurate SUVR measurements and an explicit assessment of their reliability.

Figures & tables

Figure 1: Overview of the MRI-free PET quantification and reliability assessment framework. The framework consists of four stages: CT-to-MRI synthesis, anatomical parcellation of the synthesized MRI, SUVR quantification, and conformal prediction for reliability-aware decision assessment with respect to the prespecified decision cut-off. This schematic was generated with OpenAI’s gpt-image model (version 2.0) based on the authors’ design and checked by the authors. The brain images are illustrative and do not depict study data.

Figure 2: Overview of the proposed CT-to-MRI generative model for faithful CT-to-MRI translation, which integrates multi-dataset training, data-source conditioning, and wider anatomical context within a Brownian Bridge Diffusion Model. Paired CT and T1-weighted MRI from three datasets are combined into a unified training pool. (B1) 15 consecutive brain CT slices are used to predict the central three T1-weighted MRI slices, providing broader anatomical context for anatomically faithful synthesis. (B2) Two conditioning signals are provided: the target MRI intensity histogram, which promotes slice-consistent synthesis following Choo et al. (2024) [ 7 ] , and the data-source condition, which accounts for inter-dataset heterogeneity. (B3) CT-to-MRI synthesis is performed using a Brownian Bridge Diffusion Model with a U-Net backbone, where both conditioning signals are injected into every denoising step. (B4) The model generates three consecutive T1-weighted MRI slices. The segmentation-guided training module explored during model development is omitted because it is not part of the final model. This schematic was generated with OpenAI’s gpt-image model (version 2.0) based on the authors’ design and checked by the authors. The brain images are illustrative and do not depict study data.

Figure 3: Cumulative ablation of the proposed design choices on the validation sets. Thin lines show paired per-subject performance, and the dark line with error bars denotes the mean $\pm$ SEM. H1 denotes multi-dataset training, H2 data-source conditioning, H3 expanded multi-slice anatomical context, and H4 segmentation-teacher feedback. Statistical significance was assessed between consecutive models using one-sided paired Wilcoxon signed-rank tests ( ${}^{*}p<0.05$ , ${}^{**}p<0.01$ , ${}^{***}p<0.001$ , ns not significant). Because H1–H3 progressively improved synthesis performance whereas H4 provided no significant additional benefit, the final model includes H1–H3 only.

	Severance ( $N=22$ )		OASIS-3 ( $N=117$ )
Method	PSNR	$\mathrm{SSIM}_{3\mathrm{D}}$	PSNR	$\mathrm{SSIM}_{3\mathrm{D}}$
Kim2025	$23.42\pm 0.83$	$0.874\pm 0.013$	$23.24\pm 0.68$	$0.807\pm 0.018$
LSGAN	$22.62\pm 1.19$	$0.856\pm 0.021$	$23.60\pm 0.99$	$0.840\pm 0.020$
ALDM	$21.16\pm 1.06$	$0.830\pm 0.021$	$20.95\pm 0.89$	$0.771\pm 0.021$
Choo2024	$25.14\pm 1.86$	$0.920\pm 0.019$	$23.48\pm 0.82$	$0.843\pm 0.019$
Ours	$\mathbf{26.02\pm 2.04}$	$\mathbf{0.937\pm 0.018}$	$\mathbf{24.69\pm 0.96}$	$\mathbf{0.877\pm 0.018}$

Table 1: Quantitative evaluation of CT-to-MRI synthesis. PSNR (dB,

\uparrow

) and

\mathrm{SSIM}_{3\mathrm{D}}

(

\uparrow

), mean

\pm

SD over subjects. Significance by a one-sided Wilcoxon signed-rank test against Choo2024. Best in bold , second-best underlined . Our method achieves the highest fidelity on both datasets and metrics.

Figure 4: Qualitative comparison of CT-to-MRI synthesis methods on a representative OASIS-3 subject. The columns show the CT input, ground-truth T1-weighted MRI, proposed method, and four comparison methods. The bottom row shows voxel-wise absolute error on the corresponding axial cortical slice. The proposed method shows close anatomical agreement with the ground truth, with residual errors largely localized around cortical boundaries.

Region	Metric	Kim2025	LSGAN	ALDM	Choo2024	Ours
Cerebral cortex	DSC $\uparrow$	$0.566\pm 0.020$	$0.620\pm 0.027$	$0.588\pm 0.020$	$0.655\pm 0.025$	$\mathbf{0.706\pm 0.024}$
Cerebral cortex	HD95 (mm) $\downarrow$	$3.972\pm 0.693$	$2.379\pm 0.382$	$2.141\pm 0.123$	$2.317\pm 0.211$	$\mathbf{1.677\pm 0.228}$
Cerebral WM	DSC $\uparrow$	$0.767\pm 0.013$	$0.809\pm 0.015$	$0.774\pm 0.015$	$0.817\pm 0.016$	$\mathbf{0.849\pm 0.015}$
Cerebral WM	HD95 (mm) $\downarrow$	$3.755\pm 0.358$	$3.362\pm 0.433$	$2.761\pm 0.212$	$2.804\pm 0.244$	$\mathbf{2.125\pm 0.153}$
Cerebellar GM (ref.)	DSC $\uparrow$	$0.867\pm 0.020$	$0.878\pm 0.018$	$0.873\pm 0.023$	$0.891\pm 0.019$	$\mathbf{0.904\pm 0.014}$
Cerebellar GM (ref.)	HD95 (mm) $\downarrow$	$2.755\pm 0.478$	$2.538\pm 0.371$	$2.481\pm 0.695$	$2.083\pm 0.332$	$\mathbf{1.873\pm 0.302}$

Table 2: Anatomical agreement of FreeSurfer parcellations derived from synthesized MRI with those derived from the ground-truth MRI. Results are reported as mean

\pm

SD over OASIS-3 test subjects (N = 117 for Ours and Choo2024; N = 116 for Kim2025, LSGAN, and ALDM, each excluding one subject whose FreeSurfer parcellation failed). Dice similarity coefficient (DSC,

\uparrow

) measures regional overlap, and HD95 (mm,

\downarrow

) measures boundary agreement. Best results for each region and metric are shown in bold , and second-best results are underlined .

Figure 5: Voxel-wise parcellation disagreement relative to the ground-truth MRI parcellation for the representative OASIS-3 subject shown in Fig. 4 . The leftmost panel shows the reference aparc+aseg parcellation derived from the ground-truth T1-weighted MRI. The remaining panels show disagreement for each CT-to-MRI synthesis method. Red indicates over-segmentation, blue under-segmentation, and yellow assignment to an incorrect anatomical region.

Pipeline	MAE $\downarrow$	$r$ $\uparrow$	Mean bias	95% LoA	Conc. $\uparrow$	FN/FP $\downarrow$
PET-only
PET-template (GAAIN-derived)	$0.162\pm 0.095$	0.952	$0.119^{***}$	$[-0.168,0.406]$	89.7%	0/9
PET-template (OASIS-derived)	$0.131\pm 0.080$	0.973	$0.049^{**}$	$[-0.238,0.336]$	96.6%	0/3
CT-based
Presotto2018	$0.055\pm 0.051$	0.989	$-0.032^{***}$	$[-0.165,0.101]$	97.7%	1/1
Matsuda2021	$0.053\pm 0.062$	0.982	$-0.006$	$[-0.166,0.155]$	98.9%	0/1

Table 3: Agreement between MRI-free MCSUVR estimates and the real MRI-based reference across pipelines (

N=87

). MAE is reported in SUVR units as mean

\pm

SD, and

r

denotes the Pearson correlation coefficient. Mean bias and 95% limits of agreement (LoA) are derived from Bland–Altman analysis. Subject-level differences (MRI-free minus real MRI-based MCSUVR) were tested against zero using a two-sided one-sample Wilcoxon signed-rank test with Bonferroni correction, and the resulting significance is marked on the mean bias ( ∗∗ adjusted

p<0.01

, ∗∗∗ adjusted

p<0.001

). Amyloid-positivity concordance is evaluated at the MCSUVR cut-off of 1.42, with FN and FP denoting false-negative and false-positive classifications. Best values are shown in bold and second-best values are underlined . The proposed method shows the closest overall agreement with the MRI-based reference.

Figure 6: Agreement between real MRI-based and MRI-free MCSUVR estimates for PiB PET in the OASIS-3 test set ( $N=87$ ). Results are shown for two PET-only pipelines, two CT-based pipelines (Presotto2018 and Matsuda2021), and five CT-to-MRI-based pipelines. Each point represents one subject. Crosses indicate amyloid-positivity classification disagreements with the MRI-based reference. The gray dashed line denotes the identity line, the solid line represents the least-squares fit, and the colored dashed lines mark the amyloid-positivity cut-off of 1.42. The CT-to-MRI-based pipelines, particularly the proposed method, show the closest agreement with the real MRI-based reference, whereas the PET-only pipelines show regression toward the cohort mean and the CT-based pipelines show greater scatter than the CT-to-MRI-based pipelines.

Model	Frontal	Parietal	Lateral Temporal	Precuneus	Anterior Cingulate	Posterior Cingulate
Kim2025	$0.0671\pm 0.0381$	$0.0329\pm 0.0340$	$0.0352\pm 0.0345$	$0.0402\pm 0.0432$	$0.0810\pm 0.0616$	$0.0458\pm 0.0454$
LSGAN	$0.0432\pm 0.0256$	$0.0321\pm 0.0242$	$0.0582\pm 0.0296$	$0.0328\pm 0.0344$	$0.0466\pm 0.0398$	$0.0388\pm 0.0319$
ALDM	$0.0366\pm 0.0727$	$0.0495\pm 0.1326$	$0.0421\pm 0.0801$	$0.0574\pm 0.1184$	$0.0513\pm 0.1126$	$0.0526\pm 0.0828$
Choo2024	$0.0218\pm 0.0154$	$0.0249\pm 0.0204$	$0.0248\pm 0.0202$	$0.0313\pm 0.0254$	$0.0388\pm 0.0309$	$0.0284\pm 0.0253$
Ours	$\mathbf{0.0173\pm 0.0148}$	$\mathbf{0.0159\pm 0.0148}$	$\mathbf{0.0169\pm 0.0172}$	$\mathbf{0.0207\pm 0.0229}$	$\mathbf{0.0258\pm 0.0200}$	$\mathbf{0.0234\pm 0.0215}$

Table 4: Mean absolute error (MAE) of regional SUVR estimates across six amyloid-relevant cortical regions. Results are reported as mean

\pm

SD over subjects (

N=87

). Lower values indicate better agreement with the real MRI-based reference. Best values are shown in bold .

Figure 7: Zero-shot external evaluation of CT-to-MRI synthesis and anatomical parcellation on the GARD cohort. (a) Representative CT input, paired real T1-weighted MRI, and synthetic MRIs generated by the compared methods in the coronal, sagittal, and axial planes. The bottom row shows voxel-wise absolute error relative to the real MRI for the corresponding axial slice. (b) Subject-level $\mathrm{SSIM}_{3\mathrm{D}}$ among subjects with complete synthesis results ( $N=84$ ). Thin lines connect measurements from the same subject across methods, and diamonds indicate the cohort mean. The proposed method achieves the highest $\mathrm{SSIM}_{3\mathrm{D}}$ in 62 of 84 subjects, while LSGAN achieves the highest value in the remaining 22 subjects. (c) Agreement between FreeSurfer parcellations derived from synthetic and real MRIs among subjects with valid results for all compared methods ( $N=80$ ). Points and horizontal error bars indicate mean $\pm$ SD for the Dice similarity coefficient (DSC, $\uparrow$ ) and the 95th-percentile Hausdorff distance (HD95, mm, $\downarrow$ ) across six anatomical regions.

Table 6: Empirical coverage and prediction interval width for MRI-free PET quantification using deep-ensemble-based conformal prediction. MCSUVR results are reported at target coverage levels of 80%, 90%, and 95%, and regional SUVR results at the 90% target coverage. Empirical coverage denotes the proportion of test subjects whose MRI-based reference SUVR falls within the prediction interval, with counts shown in parentheses as covered subjects/total subjects. The 95% CI column reports two-sided Wilson confidence intervals for coverage in percentage units. Interval width is reported in SUVR units as mean $\pm$ SD across test subjects ( $N=87$ ).

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Setting
Framework	PyTorch 2.6
Optimizer	Adam, lr $10^{-4}$ , bfloat16 autocast; EMA decay 0.995
Effective batch size	16 (a mini-batch of 2 with 8-step gradient accumulation on a single A100 GPU)
Synthesis input/output	15 axial CT slices $\rightarrow$ central 3 MRI slices (H3)
Diffusion backbone	BBDM, U-Net score network
Diffusion process	Brownian bridge with $T=1000$ steps, linear $m_{t}$ , maximum variance 1, the network is trained to predict $m_{t}(y-x_{0})+\sqrt{\delta_{t}}\,\epsilon$ with an $\ell_{1}$ loss

Appendix

Table A.1: Key implementation settings of the proposed pipeline.

Setting	Choo2024	ALDM	Kim2025	LSGAN
Implementation source	Released CT2MRI code [ S1 ]	Released network modules [ S2 ]	Re-implemented from the publication [ S3 ]	3D pix2pix with least-squares loss [ S4 , S5 ]
Input $\rightarrow$ output	3 CT slices $\rightarrow$ 3 MRI slices	3D CT $\rightarrow$ 3D MRI	3 CT slices $\rightarrow$ central MRI slice	3D CT $\rightarrow$ 3D MRI
Network	2D BBDM U-Net; MRI histogram condition	3D first-stage encoder/decoder with MS-SPADE; latent diffusion U-Net	2.5D residual U-Net (16/32/64 encoder channels; six residual blocks)	3D U-Net generator and 3D PatchGAN discriminator
Training objective	Brownian bridge, $\ell_{1}$ ; $T=1000$	First stage $\ell_{1}$ ; second-stage $\epsilon$ -prediction MSE, $T=1000$	$\ell_{2}$	Least-squares adversarial $+100\ell_{1}$
Optimizer	Adam, $10^{-4}$ , $\beta=(0.9,0.999)$ ; EMA 0.995	AdamW, $10^{-4}$ , weight decay $10^{-4}$ ; EMA 0.9999	Adam, $10^{-4}$ , $\beta=(0.5,0.999)$	Adam, $2\times 10^{-4}$ , $\beta=(0.5,0.999)$
Effective batch	16 (2 $\times$ 8 accumulation)	4 (1 $\times$ 4 accumulation)	16	4 (1 $\times$ 4 accumulation)

Appendix

Table A.2: Implementation settings of the four CT-to-MRI synthesis baselines in this study. Training length and checkpoint selection are reported separately for OASIS-3 and Severance where applicable. EMA: exponential moving average; ISTA: iterative shrinkage-thresholding-style sampling.

Implementation	Template or normalization settings	Target atlas / reference region
PET-only template-based normalization
PET-template (GAAIN-derived)	PiB template constructed from the public GAAIN AD-100 and YC-0 calibration PET/MRI datasets	FreeSurfer DK parcellation of the MNI152 template / GAAIN cerebellar GM
PET-template (OASIS-derived)	PiB template from 197 OASIS-3 subjects. Normalization used 8-mm source smoothing and default regularization.	Population atlas from 200 OASIS-3 FreeSurfer parcellations / cerebellar cortex
CT-based anatomical normalization
Presotto2018	HU $<-300$ set to $-1024$ . Bias correction disabled. Gaussians per tissue class were 1/1/2/2/2/2.	DK atlas from atlasreader (cvs_avg35 in MNI152) / cerebellar cortex
Matsuda2021	Skull-based initialization and CT coregistration to avg152T1. Default segmentation settings without HU thresholding.	DK atlas from atlasreader (cvs_avg35 in MNI152) / cerebellar cortex

Appendix

Table A.3: Main differences between the non-synthesis implementations. Both PET-only implementations use SPM Old Normalise. Both CT-based implementations apply CT-derived unified-segmentation deformations to PET. DK denotes Desikan–Killiany, and GM denotes grey matter.

	Severance ( $N=10$ )		OASIS-3 ( $N=20$ )
Model	PSNR $\uparrow$	$\mathrm{SSIM}_{3\mathrm{D}}$ $\uparrow$	PSNR $\uparrow$	$\mathrm{SSIM}_{3\mathrm{D}}$ $\uparrow$
BBDM [ S1 ]	$25.26\pm 1.77$	$0.912\pm 0.034$	$23.51\pm 0.48$	$0.846\pm 0.016$
$+$ H1	$25.46\pm 1.78^{\text{ns}}$	$0.919\pm 0.032^{***}$	$23.59\pm 0.48^{*}$	$0.848\pm 0.016^{*}$
$+$ H1 $+$ H2	$25.64\pm 1.86^{**}$	$0.922\pm 0.032^{***}$	$23.69\pm 0.50^{*}$	$0.852\pm 0.015^{***}$
$+$ H1 $+$ H2 $+$ H3 (proposed)	$\mathbf{26.26\pm 1.99}^{**}$	$\mathbf{0.931\pm 0.034}^{**}$	$\mathbf{24.76\pm 0.53}^{***}$	$\mathbf{0.881\pm 0.013}^{***}$
$+$ H4 (explored)	$26.16\pm 2.08^{\text{ns}}$	$0.929\pm 0.037^{\text{ns}}$	$24.37\pm 0.60^{\text{ns}}$	$0.872\pm 0.014^{\text{ns}}$

Appendix

Table A.4: Cumulative ablation analysis on the validation sets (Severance

N=10

and OASIS-3

N=20

). Results are reported as mean

\pm

SD for masked PSNR and

\mathrm{SSIM}_{3\mathrm{D}}

. Superscripts indicate one-sided paired Wilcoxon significance relative to the preceding configuration (

{}^{*}p<0.05

{}^{**}p<0.01

{}^{***}p<0.001

{}^{\text{ns}}

not significant).

Figure A.1: Effect of through-plane CT context on synthesis performance on the validation sets. The number of input axial CT slices was varied from 3 to 15 while all other model components were fixed. Thin lines show paired subject-level performance, and the dark line with error bars denotes the mean $\pm$ SEM. Statistical significance was assessed between consecutive context sizes using two-sided Wilcoxon signed-rank tests with Holm correction for multiple comparisons ( ${}^{*}p<0.05$ , ${}^{**}p<0.01$ , ${}^{***}p<0.001$ , ns not significant).

Figure A.2: Per-subject paired comparison between the proposed method and Choo2024, the strongest baseline, across both cohorts and evaluation metrics. Each thin line connects paired values from a single subject. Blue lines indicate better performance of the proposed method, whereas red lines indicate the opposite. Diamonds denote group means. Statistical significance was assessed using one-sided Wilcoxon signed-rank tests on the paired subject-level differences ( ${}^{***}p<0.001$ ).

Pipeline	Frontal	Parietal	Lat. temporal	Precuneus	Ant. cingulate	Post. cingulate
PET-template (GAAIN-derived)	$0.173\pm 0.103$	$0.160\pm 0.096$	$0.157\pm 0.094$	$0.151\pm 0.101$	$0.251\pm 0.149$	$0.170\pm 0.116$
PET-template (OASIS-derived)	$0.140\pm 0.082$	$0.146\pm 0.073$	$0.132\pm 0.074$	$0.122\pm 0.075$	$0.154\pm 0.105$	$0.140\pm 0.099$
Presotto2018	$0.098\pm 0.075$	$0.142\pm 0.082$	$0.041\pm 0.041$	$0.072\pm 0.072$	$0.139\pm 0.104$	$0.128\pm 0.126$
Matsuda2021	$0.087\pm 0.082$	$0.154\pm 0.084$	$0.052\pm 0.050$	$0.090\pm 0.081$	$0.191\pm 0.122$	$0.168\pm 0.133$
Ours	$\mathbf{0.017\pm 0.015}$	$\mathbf{0.016\pm 0.015}$	$\mathbf{0.017\pm 0.017}$	$\mathbf{0.021\pm 0.023}$	$\mathbf{0.026\pm 0.020}$	$\mathbf{0.023\pm 0.021}$

Appendix

Table A.5: MAE of regional SUVR estimates for the non-synthesis MRI-free baselines (

N=87

), reported as mean

\pm

SD in SUVR units. The proposed method is shown for reference using the results from Table 4 . Target regions of the non-synthesis pipelines were defined in MNI152 space and therefore do not coincide exactly with the native-space FreeSurfer regions of the reference.

	80% target		90% target		95% target
Region	Coverage	Width	Coverage	Width	Coverage	Width
Frontal	77.0% (67/87) [67.1, 84.6]	$0.070\pm 0.037$	88.5% (77/87) [80.1, 93.6]	$0.094\pm 0.049$	95.4% (83/87) [88.8, 98.2]	$0.120\pm 0.063$
Parietal	83.9% (73/87) [74.8, 90.2]	$0.061\pm 0.032$	88.5% (77/87) [80.1, 93.6]	$0.069\pm 0.036$	89.7% (78/87) [81.5, 94.5]	$0.074\pm 0.039$
Lateral temporal	77.0% (67/87) [67.1, 84.6]	$0.056\pm 0.029$	90.8% (79/87) [82.9, 95.3]	$0.074\pm 0.039$	93.1% (81/87) [85.8, 96.8]	$0.086\pm 0.045$
Precuneus	85.1% (74/87) [76.1, 91.1]	$0.075\pm 0.035$	92.0% (80/87) [84.3, 96.0]	$0.101\pm 0.046$	94.3% (82/87) [87.2, 97.5]	$0.116\pm 0.053$
Anterior cingulate	82.8% (72/87) [73.5, 89.3]	$0.094\pm 0.042$	96.6% (84/87) [90.3, 98.8]	$0.185\pm 0.082$	96.6% (84/87) [90.3, 98.8]	$0.193\pm 0.086$

Appendix

Table A.6: Regional conformal prediction performance across target coverage levels. Empirical coverage and prediction interval width are reported for each regional SUVR at target coverage levels of 80%, 90%, and 95%. Empirical coverage denotes the proportion of test subjects whose real MRI-based SUVR falls within the corresponding prediction interval, with counts shown in parentheses as covered subjects/total subjects. Two-sided 95% Wilson confidence intervals for coverage are shown in brackets below each coverage value, in percentage units. Interval width is reported in SUVR units as mean

\pm

SD across test subjects (

N=87

Figure A.3: Subject-level conformal prediction intervals for the six regional SUVR measures at the 90% target coverage. Within each region, subjects are ordered by the ensemble-mean MRI-free SUVR estimate. Green circles and red crosses indicate whether the corresponding MRI-based reference SUVR falls inside or outside the conformal prediction interval, respectively. Empirical coverage and mean prediction interval width for each region, including results at the 80% and 95% target coverage levels, are reported in Table A.6 .

Figure A.4: Sensitivity of amyloid-positivity decision reliability to the MCSUVR cut-off using fixed 90% conformal prediction intervals. The MCSUVR cut-off was varied from 1.30 to 1.50 while the conformal prediction intervals were held fixed. (A) Disagreement rate between the MRI-free point-estimate amyloid-positivity decision and the real MRI-based reference decision for all subjects and for subjects retaining a confident conformal decision. (B) Proportion of subjects designated indeterminate because their conformal prediction interval spanned the decision cut-off. The vertical dashed line marks the prespecified PiB PET MCSUVR cut-off of 1.42 used in the main analysis.

Cut-off	Ref. +/–	FP/FN	Confident	Indet.	Captured	Conf. disag.
1.30	24/63	0/0	84/87	3/87	–	0/84
1.35	22/65	1/1	83/87	4/87	2/2	0/83
1.40	21/66	1/0	86/87	1/87	1/1	0/86
1.42	21/66	0/0	85/87	2/87	–	0/85
1.45	20/67	0/0	85/87	2/87	–	0/85
1.50	20/67	0/0	86/87	1/87	–	0/86

Appendix

Table A.7: Amyloid-positivity decision reliability across representative MCSUVR cut-offs using fixed 90% conformal prediction intervals. MRI-free point-estimate decisions are compared with the corresponding real MRI-based reference decisions. Captured disagreements denote discordant point-estimate decisions that were designated indeterminate by the conformal decision rule. Confident disagreement denotes discordant cases remaining among subjects assigned a confident decision.

Measure	Target	OASIS-3 cal. $\rightarrow$ GARD		GARD-specific calibration
		Coverage	Width	Coverage	Width
Composite MCSUVR
MCSUVR	80%	42.7%	0.075	$81.4\pm 8.8$ %	0.160
	90%	62.2%	0.119	$90.7\pm 6.2$ %	0.203
	95%	78.0%	0.151	$95.2\pm 4.6$ %	0.252
Regional SUVR at 90% target

Appendix

Table A.8: External transfer and target-cohort recalibration of conformal prediction on GARD using the ensemble-normalized nonconformity score. For OASIS-3 calibration

\rightarrow

GARD, conformal quantiles estimated from the OASIS-3 calibration set were applied directly to the GARD cohort. For GARD-specific calibration, coverage is reported as mean

\pm

SD across 1,000 random split-half partitions, and width denotes the corresponding mean prediction interval width averaged across splits.

Medical Image Synthesis

Medical image synthesis creates, translates, reconstructs, and edits clinical images to fill missing data and support analysis.

See all 312 papers

When Brains Disagree: Biological Ambiguity Underlies the Challenge of Amyloid PET Synthesis from Structural MRI
May 12, 2026Louise E. G. Baron, Ross Callaghan, David M. Cash +3Alzheimer's Disease Medical Image Synthesis
Hawkes Institute, University College London, UK · Department of Medical Physics and Biomedical Engineering, University College London, UK · AINOSTICS Ltd, Manchester, UK +3
Bridging Modalities on the Cortex: Surface-based MRI to PET Translation with a Diffusion Bridge
Sep 17, 2026Yitong Li, Alexandra Samoylova, Fabian Bongratz +4Image-to-Image Translation Vision Transformer
Technical University of Munich (TUM), TUM University Hospital, and Munich Center for Machine Learning (MCML), Germany · Department of Neurology, TUM University Hospital, Germany · Department of Neuroradiology, TUM University Hospital, Germany +1
Cross-Modal Knowledge Distillation for PET-Free Amyloid-Beta Detection from MRI
Apr 14, 2026Francesco Chiumento, Julia Dietlmeier, Ronan P. Killeen +3Alzheimer's Disease Medical Image Classification
Dublin City University, Ireland · Insight Research Ireland Centre for Data Analytics, Ireland · St. Vincent’s University Hospital, Ireland +1

May 12, 2026J/K move · Enter open · S save

When Brains Disagree: Biological Ambiguity Underlies the Challenge of Amyloid PET Synthesis from Structural MRI

Louise E. G. Baron, Ross Callaghan, David M. Cash, Philip S. J. Weston+2

Hawkes Institute, University College London, UK · Department of Medical Physics and Biomedical Engineering, University College London, UK · AINOSTICS Ltd, Manchester, UK+3

Alzheimer's Disease Medical Image Synthesis Image Inverse Problems

Open paper AskPDF arXiv

Structural MRI-to-amyloid PET synthesis has been proposed as a non-invasive alternative for amyloid assessment in Alzheimer's disease (AD). However, reported performance of identical models varies widely across studies, and increasingly complex architectures have not led to consistent gains. This inconsistency is thought to be caused by a fundamental biological ambiguity: MRI captures neurodegeneration, while PET measures amyloid pathology - two processes that are often temporally decoupled in AD. As a result, similar MRI patterns may correspond to different amyloid states, creating ambiguous one-to-many mappings. MRI-to-amyloid PET synthesis may therefore be intrinsically ill-posed; however, this idea has yet to be tested scientifically. The aim of this work is to test this hypothesis through two controlled experiments. We first control the training distribution by stratifying paired MRI-PET data by amyloid and neurodegeneration status. Using two standard synthesis models under a controlled design, we show that biologically unambiguous mappings are learnable in isolation, but performance collapses when data ambiguity is introduced. This demonstrates that ambiguity in the data distribution, rather than architectural capacity, constrains performance. Second, we show that introducing orthogonal biological information in the form of plasma biomarkers resolves this ambiguity. When multimodal inputs are incorporated, performance improves and stability is restored. Together, these findings suggest that limited and inconsistent performance in MRI-to-amyloid PET synthesis is explained by intrinsic biological ambiguity, and that stable, meaningful progress requires multimodal integration rather than architectural complexity.

Abstract

Medical Image Synthesis

Explore similar work

When Brains Disagree: Biological Ambiguity Underlies the Challenge of Amyloid PET Synthesis from Structural MRI

Bridging Modalities on the Cortex: Surface-based MRI to PET Translation with a Diffusion Bridge

Cross-Modal Knowledge Distillation for PET-Free Amyloid-Beta Detection from MRI