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

Technical Summary: Biological Ambiguity Underlies the Challenge of Amyloid PET Synthesis from Structural MRI

Problem Statement
The synthesis of amyloid positron emission tomography (PET) images from structural magnetic resonance imaging (MRI) has been proposed as a non-invasive, scalable alternative for assessing amyloid pathology in Alzheimer's disease (AD). However, existing deep learning approaches report widely inconsistent performance, with classification accuracies ranging from near-chance levels to over 80%, even when using identical baseline models and datasets. While previous literature often attributes these limitations to architectural deficiencies or insufficient data, this paper posits a fundamental biological cause: the mapping from MRI to amyloid PET is intrinsically ill-posed.

In AD, amyloid accumulation and neurodegeneration are temporally decoupled processes. Amyloid deposition can precede neurodegeneration by years, and neurodegeneration can persist after amyloid levels stabilize. Consequently, similar structural MRI patterns (reflecting neurodegeneration) can correspond to distinct amyloid states (positive or negative), creating an ambiguous one-to-many mapping. The authors hypothesize that this biological ambiguity, rather than model capacity, is the primary constraint on synthesis performance.

Methodology
To test this hypothesis, the authors conducted two controlled experiments using 2,884 paired T1-weighted MRI and amyloid PET (AV45 tracer) scans from 1,358 subjects in the Alzheimer's Disease Neuroimaging Initiative (ADNI). Subjects were stratified based on amyloid status (A+/A–, defined by cortical SUVR ≥1.11) and neurodegeneration status (N+/N–, quantified via AVRA atrophy scores).

Two standard synthesis architectures were evaluated: a conditional Generative Adversarial Network (pix2pix) and a Latent Diffusion Model (LDM).

• Experiment I: Isolating Biological Ambiguity
  The authors manipulated the training distribution to control for ambiguity:

  1. Baseline Model: Trained on all available data (mixed A/N states).
  2. Concordant Model: Trained only on biologically aligned cases (A–/N– and A+/N+).
  3. Discordant Model: Trained only on biologically dissociated cases (A–/N+ and A+/N–).
     All models were evaluated on a fixed, held-out test set containing all biological combinations.

• Experiment II: Biological Completion
  To determine if the instability arises from missing biological information, the authors conditioned the Baseline model on plasma biomarkers (pT217, Aβ42, Aβ40, NfL, GFAP). Using Feature-wise Linear Modulation (FiLM) within the pix2pix generator, the model was retrained on a subset of 589 MRI-PET pairs with corresponding plasma data to see if orthogonal molecular signals could resolve the ambiguity.

Key Results

• Ambiguity Isolation: When trained on biologically consistent subsets (Concordant or Discordant), both architectures achieved strong performance (balanced accuracy ~82–83% and correlation ~0.51–0.54) within their respective in-distribution test sets. However, when the same architectures were trained on the full, mixed dataset (Baseline), performance collapsed (balanced accuracy dropped to ~57% for pix2pix and ~51% for LDM, with correlation near zero).
• Learning Shortcuts: The Baseline model trained on mixed data exhibited signs of converging toward an "averaged" solution, characterized by unstable performance across subsets, imbalanced sensitivity/specificity, and a compressed dynamic range of predicted amyloid loads. This suggests the model learns a shortcut to the mean rather than a biologically grounded mapping.
• Biological Completion: Introducing plasma biomarkers as conditioning variables significantly restored performance. On the same test subset, the MRI+plasma model achieved a balanced accuracy of 85.3% and a correlation of 0.75, compared to 64.5% and 0.21 for the MRI-only Baseline. Crucially, the multimodal model maintained consistent performance across both concordant and discordant subsets, indicating that the additional molecular signals successfully disentangled the ambiguous mapping.

Significance and Claims
The paper concludes that the inconsistent and limited performance observed in MRI-to-amyloid PET synthesis is not a failure of architectural complexity but a reflection of the intrinsic biological ambiguity of the task. Because MRI captures neurodegeneration while PET measures amyloid, and these processes are not strictly coupled, MRI alone cannot uniquely determine amyloid status.

The authors argue that meaningful progress in this field requires a shift in strategy:

1. Acknowledging Ill-Posedness: Recognizing that increasing model complexity or data quantity alone cannot solve a problem where the input does not uniquely determine the output.
2. Multimodal Integration: Stable and accurate synthesis requires the integration of complementary biological signals (such as plasma biomarkers) to resolve the ambiguity inherent in structural imaging.

The study provides empirical evidence that the instability in current synthesis models is data-driven rather than model-driven, suggesting that future solutions must explicitly account for biological heterogeneity through multimodal inputs.