DisQ-HNet: A Disentangled Quantized Half-UNet for Interpretable Multimodal Image Synthesis Applications to Tau-PET Synthesis from T1 and FLAIR MRI

DisQ-HNet is a novel, interpretable framework that synthesizes tau-PET images from T1 and FLAIR MRI by employing a Partial Information Decomposition-guided vector-quantized encoder and a Half-UNet decoder to disentangle modality contributions while preserving anatomical details and disease-relevant signals for Alzheimer's disease analysis.

The Gist

Imagine you are trying to predict the future of a city's traffic (Alzheimer's disease) by looking at a single, expensive, and hard-to-get satellite photo (a Tau-PET scan). This photo shows exactly where the "traffic jams" (toxic tau proteins) are clogging the brain's roads.

However, getting these satellite photos is expensive, involves radiation, and isn't available everywhere. Fortunately, we already have free, easy-to-get, high-resolution street maps (MRI scans) that show the city's layout and buildings.

The problem? Street maps don't show traffic jams. They just show the buildings.

This paper introduces a new AI tool called DisQ-HNet. Think of it as a super-smart Traffic Prediction Architect. Its job is to take two different street maps (T1 and FLAIR MRIs) and draw a brand-new, highly accurate "traffic map" (Tau-PET) that looks just like the expensive satellite photo, but without needing the satellite.

Here is how it works, using simple analogies:

1. The "Secret Decoder Ring" (Disentanglement)

Most AI tools are like a blender: they throw all the information from the two maps into a pot, mix it up, and hope the result looks right. But if the result is wrong, you have no idea why. Was it the buildings? The roads? The weather?

DisQ-HNet is different. It uses a "Secret Decoder Ring" based on a math concept called PID (Partial Information Decomposition). Instead of blending everything, it sorts the information into three distinct buckets:

• The Redundant Bucket (Shared Info): Things both maps agree on, like the shape of the brain's main highways.
• The Unique Bucket (Special Info): Things only one map sees. For example, Map A might see a specific type of tissue swelling, while Map B sees a different kind of fluid.
• The Complementary Bucket (The Magic Mix): This is the most important part. It's the "Aha!" moment that only happens when you combine the two maps. It's like realizing that because the road is narrow (Map A) and the weather is rainy (Map B), traffic will jam here. This bucket captures the complex interactions that neither map could predict alone.

By sorting the data this way, the AI doesn't just guess; it knows exactly which part of the brain's layout contributed to the prediction.

2. The "Half-UNet" (The Smart Blueprint)

Usually, AI models that draw images use "skip connections." Imagine an architect who draws a rough sketch, then skips the middle part of the process and just pastes the original blueprint onto the final drawing to make it look detailed. This is fast, but it's cheating! The AI didn't actually learn how to draw the details; it just copied them.

DisQ-HNet uses a "Half-UNet" approach. It removes the "cheating" shortcut. Instead of copying the original map, it uses Edge Cues (like tracing the outlines of the buildings) to guide the drawing.

• It forces the AI to learn the rules of how the traffic jams form based on the sorted information (the buckets above).
• It then uses the "outlines" of the brain to ensure the final drawing is sharp and detailed, without just copying the original input.

3. Why This Matters for Alzheimer's

In the real world, doctors need to know not just if a patient has Alzheimer's, but how far it has progressed (staging).

• Old AI: Might draw a pretty picture of a traffic jam, but if you ask, "Is this jam caused by a bridge collapse or a parade?", it can't answer. It's a "black box."
• DisQ-HNet: Can tell you, "This traffic jam is mostly due to the unique fluid signals from Map A, but the severity is driven by the interaction between the road shape and the fluid."

The Results

The researchers tested this on data from hundreds of patients.

• Accuracy: It created synthetic PET scans that looked almost identical to the real, expensive ones.
• Reliability: It was better at predicting the specific "stage" of Alzheimer's (Braak staging) than other models.
• Trust: Because it sorted the information, doctors can trust why the AI made a prediction. It's not magic; it's a transparent process.

The Bottom Line

DisQ-HNet is like a master chef who doesn't just throw ingredients into a pot. Instead, they separate the spices, the meats, and the vegetables, understand how they interact, and then cook a dish that tastes exactly like a rare, expensive meal, using only common ingredients.

This allows doctors to get the critical "traffic jam" data (Tau-PET) they need to treat Alzheimer's, using the cheap, safe, and available "street maps" (MRI) they already have, all while understanding exactly how the prediction was made.

Technical Summary

1. Problem Statement

Context: Tau positron emission tomography (tau-PET) is a critical biomarker for staging Alzheimer's Disease (AD) and tracking tau pathology (Braak staging). However, its clinical adoption is limited by high costs, radiation exposure, and logistical constraints.
Challenge: While MRI (specifically T1-weighted and FLAIR) is widely available, synthesizing tau-PET from MRI using deep learning faces two major hurdles:

1. Black-box Nature: Standard synthesis models (e.g., UNets) often act as black boxes, obscuring which input modality drives specific prediction features. This lack of interpretability hinders clinical trust.
2. Signal Entanglement: Standard architectures use skip connections that bypass the latent bottleneck, allowing low-level anatomical details to leak directly to the decoder. This prevents the model from learning a structured, disentangled representation of redundant (shared), unique (modality-specific), and complementary (synergistic) information.
3. Clinical Fidelity vs. Downstream Utility: High visual fidelity (low pixel error) does not guarantee that the synthesized PET preserves the specific disease-relevant signals required for accurate Braak staging or tau localization.

2. Methodology: DisQ-HNet

The authors propose DisQ-HNet, a framework designed to synthesize tau-PET from paired T1 and FLAIR MRI while explicitly disentangling multimodal information. The architecture consists of two main innovations:

A. PID-Guided Disentangled Quantization (DisQ-VAE)

Instead of a continuous latent space, the model uses a Vector Quantized (VQ) encoder guided by Partial Information Decomposition (PID).

• Latent Factorization: The latent space is explicitly partitioned into three components based on PID theory:
  • Redundant ($R$): Information shared by both T1 and FLAIR (anatomical structure).
  • Unique ($U_{T1}, U_{FLAIR}$): Information specific to a single modality.
  • Complementary ($C$): Synergistic information that only emerges when modalities are combined.
• Architecture:
  • An RU Encoder processes each modality to extract Redundant and Unique features.
  • A Complementary Encoder processes concatenated inputs to extract $C$.
  • Quantization: Latents are discretized using separate codebooks ($K=4096$) for Redundant/Unique and Complementary streams.
• Loss Functions:
  • PID-Inspired Mutual Information Loss: Enforces independence between unique and redundant/complementary factors and encourages agreement between redundant factors across modalities.
  • Variance Floor: Prevents "posterior collapse" (where latent variables become constant) by penalizing low-variance channels.
  • Symmetry Regularization: Random modality swapping and mixing during training ensure the redundant pathway remains modality-agnostic.

B. Edge-Conditioned Half-UNet Decoder

To preserve structural detail without reintroducing uninterpretable shortcuts:

• No Direct Skip Connections: The decoder does not receive direct feature maps from the encoder.
• Pseudo-Skip Connections: Structural guidance is injected via a structural edge pyramid derived from the input MRI (using 3D Sobel filters).
• Mechanism: The decoder is initialized with the quantized latent factors. At subsequent upsampling stages, it fuses upsampled features with learned projections of the edge pyramid.
• Consistency Loss: An auxiliary loss ensures the pseudo-skip signals remain structurally aligned with the input anatomy.

3. Key Contributions

1. Architectural Interpretability: The first framework to integrate PID theory directly into the latent bottleneck of a medical image synthesis model, explicitly separating redundant, unique, and synergistic signals.
2. DisQ-HNet Framework: A novel "Half-UNet" design that replaces standard skip connections with edge-conditioned pseudo-skips, forcing the decoder to rely on the learned latent factors rather than bypassing them.
3. Quantized Multimodal Synthesis: Demonstrates that vector quantization combined with PID constraints yields better disease-relevant signal preservation than continuous VAEs or standard UNets.
4. Shapley Attribution: Provides a mechanistic explanation of the synthesis process using Shapley values to quantify the marginal contribution of each PID component (Redundant, Unique, Complementary) to the final output.

4. Experimental Results

The model was evaluated on 605 subjects from ADNI-3 and OASIS-3 datasets, synthesizing tau-PET from T1 and FLAIR MRI.

• Reconstruction Fidelity:
  • DQ2H-MSE-Inf (the proposed model) achieved the best raw PET fidelity (MAE: 0.1043, PSNR: 18.53) and competitive SUVR performance.
  • It outperformed standard UNets and VQ-VAEs in preserving high-uptake regions (Dice score: 0.539 vs. 0.161 for UNet-MSE).
• Downstream Clinical Utility (Braak Staging):
  • DQ2H-MSE-Inf achieved the highest Hard Accuracy (0.482) and Soft Accuracy (0.779) for Braak staging, significantly outperforming baselines (e.g., UNet-MSE had ~0.096 hard accuracy).
  • It showed the strongest correlation with regional SUVR patterns in early Braak stages (I-III), whereas baselines often showed near-zero or negative correlations.
• Bias and Calibration:
  • Standard UNets exhibited significant systematic underestimation of SUVR (Bias: -18.4%).
  • DisQ-HNet reduced this bias to -3.3%, indicating better calibration for clinical thresholds.
• PID Analysis (Shapley Values):
  • Complementary information was identified as the dominant contributor to PET-specific fidelity and perceptual similarity.
  • Redundant information provided structural scaffolding.
  • Unique T1 information contributed minimally or negatively to PET-specific similarity, confirming its role is primarily anatomical rather than pathological.

5. Significance

• Clinical Translation: The paper demonstrates that interpretability should be an architectural feature, not an afterthought. By disentangling signals, the model produces syntheses that are not just visually plausible but clinically actionable for AD staging.
• Multimodal Insight: The results validate that the "synergy" between T1 and FLAIR (Complementary component) is crucial for capturing tau pathology, a finding that standard black-box models obscure.
• Generalizability: While applied to AD, the DisQ-HNet framework addresses a general class of multimodal synthesis problems where understanding the source of generated features is critical for trust and safety.

Conclusion: DisQ-HNet successfully bridges the gap between high-fidelity image synthesis and clinical interpretability, offering a robust, bias-reduced method for generating tau-PET from routine MRI, thereby potentially lowering the barrier for early Alzheimer's detection and staging.