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 · 2026

Read Cross-Modal Knowledge Distillation for PET-Free Amyloid-Beta Detection from MRI by Francesco Chiumento, Julia Dietlmeier, Ronan P. Killeen, Kathleen M. Curran, Noel E. O’Connor, Mingming Liu (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-2604-12574.txt.

Introduction

Francesco Chiumento and five colleagues predict amyloid-beta positivity from MRI alone. Amyloid positivity marks early Alzheimer's disease but usually needs costly PET scans. Their teacher model learns PET-MRI alignment with cross-modal attention. An MRI-only student then copies the teacher at two levels. Tests span four MRI contrasts and two independent datasets. Knowledge transfer works well, with best AUC reaching 0.74. Saliency maps show predictions focus on relevant cortical regions. PET-free screening gains clinical plausibility.

Catalog record

CREATORFrancesco Chiumento, Julia Dietlmeier, Ronan P. Killeen, Kathleen M. Curran, Noel E. O’Connor, Mingming Liu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59870
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.12574

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

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 Abstract Detecting amyloid- β \beta (A β \beta ) 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 \beta 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 \beta screening. Code is available at github.com/FrancescoChiumento/pet-guided-mri-amyloid-detection . 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. 1 Introduction Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that represents the predominant form of dementia. 55.2 million individuals are currently affected worldwide, with projections reaching 78 million by 2030 and costs projected to reach up to $2.8 trillion 57 ; 36 . The pathological hallmarks of AD are neurofibrillary tangles (NFTs), formed by hyperphosphorylated tau protein within neurons, and extracellular plaques of accumulated A \beta peptide accompanied by neuroinflammation, synaptic dysfunction, mitochondrial and vascular abnormalities 59 ; 45 . Currently, there is no effective cure for AD, and once it manifests, it follows an irreversible progression 59 ; 17 . An early event in AD pathophysiology is the deposition of A \beta plaques in the brain, which can occur {\sim}20 years before the onset of symptoms, offering a potential window for secondary prevention 8 . Typically, A \beta is assessed through invasive procedures such as positron emission tomography (PET) or cerebrospinal fluid (CSF) assays. PET uses radiotracers targeting amyloid or tau pathology, creating a spatial map of A \beta distribution in the brain and revealing the extent of AD pathology 4 . Although amyloid levels can be measured in individuals using PET imaging with amyloid-sensitive ligands such as Pittsburgh Compound B (PiB; {}^{11}\mathrm{C} -PiB) or florbetapir ( {}^{18}\mathrm{F} -AV-45) 58 ; 7 , amyloid-PET is expensive, not widely available, and invasive, exposing the patient to ionizing radiation. For this reason, amyloid-PET is not considered cost-effective for diagnostic use at the MCI stage 28 , yet demand is projected to increase up to {\sim} 20-fold following approval of anti-A \beta therapies (lecanemab, donanemab) that require confirmation of amyloid pathology prior to treatment initiation 54 ; 52 ; 53 . MRI, conversely, is non-invasive, widely accessible, and provides detailed structural brain information without radioactive tracers 56 . While MRI cannot directly visualize amyloid, A \beta accumulation leads to widespread brain cell loss, which manifests as atrophy in T1-weighted (T1w) sequences 9 . Several studies classify A \beta positivity (A \beta + vs. A \beta -) from PET images using deep learning models. A smaller subset uses MRI, but typically includes additional non-imaging covariates such as APOE \varepsilon 4 genotype, cognitive scores, hippocampal volumetry, or clinical diagnosis 12 ; 24 ; 30 ; 56 ; however, these covariates exhibit high missingness in practice: routine APOE testing is not recommended by ACMG guidelines 16 and is missing in 26% of NACC participants 39 ; cognitive measures are absent in 89% of EHR dementia records 35 ; and only 23% of European centres perform volumetry analysis 55 . Few studies, on the other hand, have attempted to predict amyloid status using only MRI data without other clinical variables 23 . In this work, we address binary A \beta + prediction using only MRI as the single modality. Specifically, our research question is how to effectively transfer the knowledge from a multimodal contrastive PET and MRI teacher to an MRI-only student . With this in mind, the key novelty of our work is that we propose a PET-guided knowledge distillation framework based on BiomedCLIP, a ViT pre-trained on 15M biomedical image-text pairs 60 which, to the best of our knowledge, has not been applied in this specific context. The cross-modal teacher integrates PET (query) and MRI (key) via multi-head attention (MHA) and is supervised by a combination of classification and triplet loss ; the MRI-only student learns from the teacher through feature-level and logit-level distillation with temperature annealing. At inference time, only MRI is required. Fig. Cross-Modal Knowledge Distillation for PET-Free Amyloid-Beta Detection from MRI illustrates our three-phase distillation approach. The main contributions of our work are outlined below. • PET-to-MRI knowledge distillation framework: We propose a novel cross-modal distillation approach for amyloid prediction, where an MRI-only student learns from a PET+MRI teacher via \ell_{2} -normalized feature matching and temperature-scaled logit distillation, enabling PET-free inference without non-imaging covariates. Unlike prior work using direct supervision from binary PET labels or CSF values, our teacher learns from PET’s spatial standardized uptake value ratio (SUVR) distribution 25 through cross-modal attention, transferring spatially-informed knowledge to the student model. • Contrastive teacher training: We propose a new adaptation of the triplet loss with Centiloid (CL, a standardized PET amyloid quantification scale 25 ; 20 )-aware negative mining, where online hard-negative selection ensures biochemically distinct triplets, while cross-modal attention and self-attention synthesize patient-level embeddings for triplet learning. • Multi-contrast evaluation and cross-dataset transfer: We evaluate T1w, T2w, FLAIR, and T2* sequences and their combinations on OASIS-3 and ADNI, achieving AUCs up to 0.74 and improving T1w MRI-only results by +19.7% (0.73 vs. 0.61) over Kim et al. 23 , and validate cross-dataset knowledge distillation (CDKD) 1 ; 27 . 2 Related Work 2.1 MRI-based Amyloid Prediction MRI-only methods for predicting A \beta + remain limited and achieve only moderate discriminative performance. Most studies focus on binary classification of amyloid- \beta status (A \beta + vs. A \beta -) in AD research, often using PET or CSF labels within deep learning models 12 . A smaller number of works have investigated MRI-only prediction. Kim et al. 23 trained a 3D EfficientNet on 4,056 exams (ADNI, OASIS-3, and A4), reporting AUCs of 0.61 for T1w and 0.67 for T1w+FLAIR, without clinical variables. In contrast, when structured variables are added, performance increases: Lew et al. 30 combined MRI with demographic, APOE, and cognitive information to predict PET-derived Amyloid/Tau/Neurodegeneration (ATN) status, obtaining an amyloid AUC of 0.79 using SUVR thresholded via Gaussian mixture modeling. Most prior work either relies on direct supervision or clinical variables, with few truly MRI-only pipelines and no exploitation of PET–MRI complementarity during training 23 . We address this by supervising MRI features with a PET-guided teacher through cross-modal distillation. Although structured covariates can improve prediction, their use reduces deployability in practice, since key variables (e.g., APOE genotype and detailed cognitive scores)

Read the complete text