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

Francesco Chiumento; Julia Dietlmeier; Ronan P. Killeen; Kathleen M. Curran; Noel E. O'Connor; Mingming Liu

cs.CVApr 14, 2026

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

Authors: Francesco Chiumento, Julia Dietlmeier, Ronan P. Killeen, Kathleen M. Curran, Noel E. O'Connor, Mingming Liu

Organizations: Dublin City University, Ireland · Insight Research Ireland Centre for Data Analytics, Ireland · St. Vincent’s University Hospital, Ireland · University College Dublin, Ireland

Abstract

Detecting amyloid- $β$ (A $β$ ) positivity is crucial for early diagnosis of Alzheimer's disease but typically requires PET imaging, which is costly, invasive, and not widely accessible, limiting its use for population-level screening. We address this gap by proposing a PET-guided knowledge distillation framework that enables A $β$ prediction from MRI alone, without requiring non-imaging clinical covariates or PET at inference. Our approach employs a BiomedCLIP-based teacher model that learns PET-MRI alignment via cross-modal attention and triplet contrastive learning with PET-informed (Centiloid-aware) online negative sampling. An MRI-only student then mimics the teacher via feature-level and logit-level distillation. Evaluated across four MRI contrasts (T1w, T2w, FLAIR, T2*) and two independent datasets, our approach demonstrates effective knowledge transfer (best AUC: 0.74 on OASIS-3, 0.68 on ADNI) while maintaining interpretability and eliminating the need for clinical variables. Saliency analysis confirms that predictions focus on anatomically relevant cortical regions, supporting the clinical viability of PET-free A $β$ screening. Code is available at https://github.com/FrancescoChiumento/pet-guided-mri-amyloid-detection.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of our PET-guided knowledge distillation framework. A teacher model learns cross-modal alignment between PET and MRI, guided by Centiloid-aware triplet mining. The teacher’s knowledge is then distilled to an MRI-only student for PET-free inference.

Cohort (MRI)	Sessions			Subjects			%Pos (Train)
	Train	Val	Test	Train	Val	Test
OASIS-3 (T1w $\cap$ T2w)	518	187	166	346	118	127	21.8
OASIS-3 (FLAIR $\cap$ T2*)	288	96	89	287	93	88	34.4
ADNI (T1w $\cap$ T2w)	303	102	90	212	72	68	44.9
ADNI (FLAIR $\cap$ T2*)	608	193	190	405	132	132	50.2

Table 2: Cohorts used. Splits are subject-wise.

Figure 2: Phase 1 — Teacher pre-training with binary classification loss (BCE).

Figure 3: Phase 2 — Teacher refinement with online CL-aware triplet mining.

Figure 4: Phase 3 — Knowledge distillation . MRI-only student is trained via

\ell_{2}

feature matching and temperature-scaled logit distillation; PET/CL are used only for teacher supervision.

	OASIS-3											ADNI
Modality	Sel.	@0.5				@ $\theta^{*}$				AUC	NPV	Sel.	@0.5				@ $\theta^{*}$				AUC	NPV
Modality	Sel.	F1	Acc	Prec	Rec	F1	Acc	Prec	Rec			Sel.	F1	Acc	Prec	Rec	F1	Acc	Prec	Rec
T1w & T2w available
T1w	F1	0.53	0.71	0.57	0.49	0.59	0.71	0.56	0.64	0.73	0.81	F1	0.58	0.56	0.48	0.74	0.61	0.50	0.46	0.92	0.66	0.77
T1w	Sim	0.46	0.69	0.54	0.40	0.53	0.49	0.38	0.87	0.70	0.83	Sim	0.57	0.54	0.47	0.71	0.59	0.51	0.46	0.82	0.64	0.68
T2w	F1	0.54	0.54	0.40	0.80	0.55	0.63	0.46	0.69	0.70	0.80	F1	0.51	0.63	0.59	0.45	0.57	0.51	0.45	0.76	0.62	0.65

Table 3: Test results on OASIS-3 and ADNI. Sel.: checkpoint by val. F1 or embedding similarity. Metrics:

\theta=0.5

vs. optimized

\theta^{*}

Panel of Figure 5 — Figure 5: ROC curves per sequence. (a) OASIS-3; (b) ADNI.

Method	MRI Input	Ground Truth	Clin.	Training	N	AUC [95% CI]
Kim et al. (2025) [ 23 ] (T1w)	T1w	PET (CL)	No	Direct	779	0.61 [0.59–0.63]
Kim et al. (2025) [ 23 ] (T1w+FLAIR)	T1w+FLAIR	PET (CL)	No	Direct	779	0.67 [0.65–0.70]
Wang et al. (2025) [ 56 ]	T1w	PET (SUVR ${>}$ 1.11)	Yes †	Transfer	241	0.74
Dolci et al. (2025) [ 12 ]	3-modal	CSF (A $\beta$ 42) ‡	No	Direct	–	–
Chiumento et al. (2025) [ 10 ]	T1w	PET (CL)	No	Direct	–	–
Ours T1w	T1w	PET (CL)	No	KD+Trip	166	0.73 [0.66–0.82]

Table 4: Comparison with MRI-based amyloid prediction . MRI-only inference prioritized; covariate-based method noted.

Figure 6: Saliency/HiResCAM. OASIS-3, epochs $1\to 25$ .

ADNI (Teacher) KD $\rightarrow$ OASIS-3 (Student) — Sel.=F1
	@0.5				@ $\theta^{*}$				Overall
Modality	F1	Acc	Prec	Rec	F1	Acc	Prec	Rec	AUC	NPV
T1w	0.27	0.68	0.53	0.18	0.58	0.70	0.54	0.64	0.73	0.80
T2w	0.55	0.70	0.54	0.56	0.54	0.69	0.54	0.55	0.72	0.77
T1w+T2w $\rightarrow$ T1w	0.55	0.58	0.42	0.76	0.61	0.70	0.53	0.71	0.72	0.83
T1w+T2w $\rightarrow$ T2w	0.39	0.72	0.68	0.27	0.57	0.61	0.45	0.76	0.73	0.82

Table 5: Cross-dataset knowledge distillation (CDKD). Source teacher distills target MRI-only student. Metrics at

\theta=0.5

\theta^{*}

No Pre-training (Phase 1)
		@0.5				@ $\theta^{*}$				Overall
Dataset	Modality	F1	Acc	Prec	Rec	F1	Acc	Prec	Rec	AUC	NPV
OASIS-3	T1w	0.57	0.62	0.46	0.75	0.53	0.71	0.56	0.51	0.74	0.77
OASIS-3	T2w	0.16	0.68	0.56	0.09	0.55	0.62	0.45	0.71	0.67	0.80
ADNI	T1w	0.57	0.57	0.49	0.68	0.59	0.57	0.49	0.73	0.63	0.70
ADNI	T2w	0.51	0.57	0.49	0.53	0.55	0.50	0.44	0.74	0.61	0.63

Table 6: Ablation Study. T1w/T2w phase impact (

\theta=0.5

\theta^{*}

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Abbr.	Full Form	Abbr.	Full Form
Medical & Imaging
A $\beta$	Amyloid- $\beta$	OASIS-3	Open Access Series of Imaging Studies 3 [ 27 ]
AD	Alzheimer’s disease	PET	Positron Emission Tomography
ADNI	Alzheimer’s Disease Neuroimaging Initiative [ 1 ]	PiB	Pittsburgh Compound B [ 58 ]
CL	Centiloid	PUP	PET Unified Pipeline [ 49 ]
CSF	Cerebrospinal Fluid	SUVR	Standardized Uptake Value Ratio

Appendix

Table 7: Abbreviations used throughout the main paper and supplementary material.

	Phases 1–2 (Teacher)	Phase 3 (Student)
Projection head $(p_{1},p_{2})$	$(0.5,\ 0.4)$	$(0.3,\ 0.2)$
Classification head $p$	$0.6$	$0.4$

Appendix

Table 8: Dropout schedules. Dropout rates for projection and classification heads in teacher (Phases 1–2) and student (Phase 3).

Panel of Figure 7 — Figure 7: Single vs Multi-Sequence Performance Comparison. Spider charts comparing single-sequence and multi-sequence distilled models across metrics (F1, Accuracy, Precision, Recall, AUC). Top: T1w+T2w; bottom: FLAIR+T2*. Left: OASIS-3; right: ADNI. Multi-sequence distillation improves recall.

	OASIS-3											ADNI
Sequence	$\Delta_{\text{CL}}^{\min}$	@0.5				@ $\theta^{*}$				AUC	NPV	$\Delta_{\text{CL}}^{\min}$	@0.5				@ $\theta^{*}$				AUC	NPV
Sequence	$\Delta_{\text{CL}}^{\min}$	F1	Acc	Prec	Rec	F1	Acc	Prec	Rec			$\Delta_{\text{CL}}^{\min}$	F1	Acc	Prec	Rec	F1	Acc	Prec	Rec
T1-weighted MRI
T1w	0.0	0.53	0.49	0.38	0.87	0.53	0.68	0.51	0.56	0.72	0.77	0.0	0.60	0.63	0.56	0.66	0.60	0.60	0.52	0.71	0.65	0.71
T1w	5.0	0.53	0.71	0.57	0.49	0.59	0.71	0.56	0.64	0.73	0.81	5.0	0.58	0.56	0.48	0.74	0.61	0.50	0.46	0.92	0.66	0.77
T1w	10.0	0.53	0.65	0.47	0.60	0.52	0.63	0.46	0.60	0.68	0.77	10.0	0.59	0.60	0.52	0.68	0.58	0.53	0.47	0.76	0.65	0.68

Appendix

Table 9: Ablation study: impact of Centiloid-based negative mining. Test performance on OASIS-3 and ADNI for T1w and T2w models when varying the minimum Centiloid gap

\Delta_{\text{CL}}^{\min}

between anchor and negative samples. We report metrics both at a fixed decision threshold of

0.5

and at the validation-optimized threshold

\theta^{*}

Panel of Figure 8 — Figure 8: ROC curve evolution across datasets and sequences. Validation performance at Epoch 1 (initialization) vs. Final Epoch (convergence). Top row: OASIS-3 dataset for T1w+T2w training (left) and FLAIR+T2* training (right). Bottom row: ADNI dataset with same training configurations. All models show substantial AUC improvements demonstrating effective knowledge distillation.

Panel of Figure 9 — Figure 9: Single-sequence training: attention evolution across datasets. Training progression (epochs 1, 8, 25) for models trained on individual MRI contrasts. Each row displays PET reference, target MRI, gradient saliency, and HiResCAM maps. The network progressively focuses on anatomically relevant brain structures, with consistent patterns across OASIS-3 and ADNI datasets demonstrating robust generalization.

Panel of Figure 10 — Figure 10: Multi-sequence training with single-sequence inference. Saliency/HiResCAM evolution (epochs 1, 8, 25) for models trained on paired sequences (T1w+T2w, FLAIR+T2*) and tested on individual contrasts, showing consistent spatial attention patterns across modalities.

Explore similar work

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
Evidence Before Accuracy: A MRI-PET Fusion Network for Alzheimer Disease Classification with Causal Regional Validation
Sep 28, 2026Saeid Firouzi Daghigh, Saeed AyatAlzheimer's Disease Classification Explainable Medical Image Analysis
Department of Computer Engineering and Information Technology, Payame Noor University, Tehran, Iran
Translating MRI to PET through Conditional Diffusion Models with Enhanced Pathology Awareness
Mar 19, 2026Yitong Li, Igor Yakushev, Dennis M. Hedderich +1Medical Image Synthesis Image-to-Image Translation
Lab for Artificial Intelligence in Medical Imaging, Institute for Diagnostic and Interventional Radiology, School of Medicine and Health, TUM Klinikum, Technical University of Munich (TUM), Munich, 81675, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Department of Nuclear Medicine, School of Medicine and Health, TUM Klinikum, Munich, 81675, Germany +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

Explore similar work

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

Evidence Before Accuracy: A MRI-PET Fusion Network for Alzheimer Disease Classification with Causal Regional Validation

Translating MRI to PET through Conditional Diffusion Models with Enhanced Pathology Awareness