💻 computer science

Parcel2Progression: An Anatomy-aware Longitudinal Framework for Alzheimer's Disease Diagnosis

The paper introduces Parcel2Progression (P2P), an anatomy-aware Longitudinal Transformer framework that tokenizes high-resolution 4D sMRI scans into anatomically grounded parcels to efficiently integrate variable-length longitudinal visits, achieving state-of-the-art performance in Alzheimer's disease diagnosis and progression prediction while offering improved computational scalability and clinical interpretability.

Original authors: Madhumitha Venkatesh, Shanawaj S Madarkar, Konda Reddy Mopuri

Published 2026-08-11

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Madhumitha Venkatesh, Shanawaj S Madarkar, Konda Reddy Mopuri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a bustling, ancient city. For decades, scientists have been trying to understand what happens when this city starts to crumble due to Alzheimer's disease. The trouble is, the damage doesn't happen all at once; it's a slow, quiet erosion. In the early stages, the "buildings" (brain cells) in specific neighborhoods start to shrink, but the changes are so subtle that a quick glance at a single photo of the city often misses the signs. This is where medical imaging comes in. Doctors use special cameras called MRI scanners to take 3D pictures of the brain. Usually, they look at just one picture to make a diagnosis. But because the disease is a story that unfolds over years, looking at a single snapshot is like trying to understand a movie by watching only one frame. The real clues are in the progression—how the city changes from one visit to the next. However, trying to analyze a whole movie of high-definition brain scans is incredibly hard for computers; it's like trying to read every word in a library of books simultaneously, which often forces computers to either blur the details or skip most of the pages.

This paper introduces a clever new system called Parcel2Progression (P2P) that solves this puzzle by acting like a super-smart, organized librarian. Instead of trying to read the entire library at once, P2P first divides the brain into distinct neighborhoods called "parcels" (using a standard map of the brain). It studies each neighborhood individually to understand its unique shape and health. Then, instead of just looking at one photo, it stitches together a timeline of these neighborhoods for each patient, accounting for how much time has passed between visits. The result is a model that can spot the tiny, early signs of Alzheimer's and predict who might get worse, all while keeping the computer's workload manageable. The authors found that by looking at these changing neighborhoods over time, their system became significantly better at diagnosing the disease and predicting its future than previous methods that only looked at single snapshots or tried to process the whole brain at once.

The Problem: The "Too Much Data" Dilemma

Think of the human brain as a giant, intricate city made of billions of tiny bricks. When Alzheimer's strikes, it doesn't knock down the whole city at once. It starts by quietly eroding specific districts, like the hippocampus (the memory center) or the amygdala (the emotion center). In the early stages, known as Mild Cognitive Impairment (MCI), these changes are so faint that they are almost invisible.

For a long time, computer models trying to detect this disease faced a tough choice, like a photographer with a broken camera. They could either:

Zoom out: Look at the whole brain at once, but lose the fine details needed to see the early cracks.
Zoom in: Look at the details, but only on a single day, missing the story of how the damage spreads over years.
Try to do both: Attempt to analyze high-resolution brain scans from multiple years at once. But this is computationally impossible for most computers because the amount of data grows explosively. It's like trying to watch a 4K movie of the entire city for every single day of a patient's life; the computer gets overwhelmed and crashes.

The Solution: The "Neighborhood Watch" System

The authors of this paper, Madhumitha Venkatesh and her team, built a new framework called Parcel2Progression (P2P). Imagine the brain not as a single blob, but as a map divided into 69 distinct neighborhoods (called "parcels"), based on a standard brain atlas.

Step 1: The Neighborhood Scout (Stage A)
Instead of looking at the whole brain at once, P2P sends a scout into each of the 69 neighborhoods. This scout takes a high-resolution 3D picture of just that specific district and creates a detailed report (an "embedding") on its health. Crucially, the scout does this for every neighborhood independently. This prevents the computer from getting confused by mixing up signals from different parts of the brain. It's like having 69 specialized detectives, each focusing on one block of the city, rather than one detective trying to watch the whole city at once.

Step 2: The Time-Traveling Analyst (Stage B)
Once the scouts have their reports, the second part of the system kicks in. This is the "Longitudinal Transformer." It takes the reports from a patient's different visits over the years and stitches them together. It knows that a visit today is different from a visit five years ago, so it uses the patient's age to understand the timeline. It asks: "How did the hippocampus neighborhood change between visit 1 and visit 5?"

This two-step process is the magic trick. By breaking the brain into neighborhoods first, the computer doesn't have to do the impossible math of comparing every single pixel of every scan to every other pixel. Instead, it compares the "health reports" of the neighborhoods. This makes the system fast enough to handle long histories of scans without needing a supercomputer the size of a building.

What They Found: Seeing the Unseen

The team tested P2P on data from thousands of patients from three major studies (ADNI, AIBL, and MIRIAD). They compared their new system against the best existing methods.

1. Better at Spotting the Disease
When trying to tell the difference between healthy brains and those with Alzheimer's, P2P was more accurate than any previous method. But the real win came when they looked at the "MCI" group—people who are on the edge.

The Result: By using the timeline of scans, P2P improved its ability to predict who would convert from "Mild Cognitive Impairment" to full-blown Alzheimer's by up to 7.59% in balanced accuracy compared to single-scan methods.
The Analogy: It's like being able to tell if a tree is sick just by looking at one leaf (old method) versus watching the tree lose its leaves over a whole season (P2P). The seasonal view gives a much clearer answer.

2. Handling Real-World Messiness
In the real world, patients don't visit the doctor on a perfect schedule. Some come every year, some every two years, and some skip a visit. Many computer models break if the timing isn't perfect. P2P, however, is flexible. It can take a patient's history, no matter how irregular the visits are, and still make a solid prediction.

3. The "Why" Behind the "What"
One of the coolest parts of P2P is that it doesn't just give a "Yes/No" answer; it explains why. Because it looks at neighborhoods, it can point to the map and say, "The model is worried because the hippocampus and the temporal lobe are shrinking faster than usual."

The Proof: The team tested this on a fake, synthetic brain dataset where they knew exactly which "neighborhoods" were damaged. P2P correctly identified those specific damaged areas, proving it wasn't just guessing but actually looking at the right places.

The Takeaway

The paper suggests that the key to understanding Alzheimer's isn't just looking harder at a single picture, but looking smarter over time. By treating the brain as a collection of distinct, evolving neighborhoods rather than a single static object, P2P manages to be both highly accurate and computationally efficient.

The authors found that using these long-term, neighborhood-specific views boosted performance significantly. For the hardest task—predicting who will get Alzheimer's from early warning signs—their model showed a clear advantage over older methods. While they note that the system currently works best with one type of MRI scan (T1w) and a specific brain map, the results suggest that this "neighborhood watch" approach could be a powerful tool for doctors to catch the disease earlier and understand how it moves through the brain. It turns a blurry, overwhelming movie of brain changes into a clear, chapter-by-chapter story that computers can actually read.

=== SUMMARY ===

Technical Summary: Parcel2Progression (P2P)

1. Problem Statement

Alzheimer's disease (AD) progression is a longitudinal process characterized by subtle pathological cues in early stages, particularly in Mild Cognitive Impairment (MCI). Existing computational models for neuroimaging face a fundamental bottleneck: they must either compromise spatial resolution to handle temporal sequences or limit the number of longitudinal scans to maintain spatial fidelity.

Current approaches typically fall into two categories with significant limitations:

3D Single-scan models: While they preserve volumetric context, they are structurally cross-sectional and fail to capture the spatiotemporal patterns of neurodegeneration required to differentiate stable MCI (sMCI) from progressive MCI (pMCI).
4D Longitudinal models: Existing methods often rely on aggressive downsampling (e.g., 64×64×64) or fixed-interval assumptions to manage computational costs. Others compress 3D visits into single embeddings before temporal aggregation, which collapses regional structure and weakens sensitivity to subtle, localized progression. Furthermore, many models cannot handle the irregular visit intervals and variable-length patient histories found in real-world clinical data.

The core challenge is to build a model that simultaneously captures high-resolution, fine-grained anatomical cues and arbitrary-length longitudinal histories without incurring prohibitive computational costs (specifically the quadratic scaling of global attention in 4D Vision Transformers).

2. Methodology: Parcel2Progression (P2P)

The authors propose Parcel2Progression (P2P), a Longitudinal Transformer Framework designed to unify anatomical grounding with efficient temporal modeling. The framework operates in two distinct stages:

Stage A: Atlas-guided Parcel Encoding (P2P-SA)

This stage encodes a single high-resolution 3D T1w sMRI scan into anatomically grounded representations.

Tokenization: The scan is partitioned into non-overlapping cubic patches. A shared tokenizer (MLP or DenseNet) projects these into embeddings, augmented with 3D spatial coordinates.
Parcel Grouping: Using a probabilistic brain atlas (Harvard-Oxford), the patches are assigned to $P$ anatomical parcels (e.g., hippocampus, amygdala) via soft membership. This creates parcel-specific token sets.
Local Attention: To prevent information leakage and "collapse" (where distinct regions merge into a generic representation), self-attention is performed independently within each parcel set using a shared Transformer encoder. This ensures region-specific contextualization.
Aggregation: Tokens within each parcel are aggregated using Generalized Mean (GeM) pooling to produce a single embedding per parcel, preserving the importance of small but critical structures.
Parcel Scoring: A learnable scorer assigns diagnostic importance weights to each parcel embedding. These weights are used to form a convex combination of parcel embeddings, yielding a discriminative scan-level representation.
Regularization: A "parcel-dropout" mechanism is introduced during training to prevent over-reliance on a small subset of regions.

Stage B: Age-aware Longitudinal Transformer

This stage aggregates evidence over variable-length longitudinal histories using the frozen parcel embeddings from Stage A.

Input: The model accepts a sequence of visits with irregular intervals and varying lengths. Patient age is injected as an additive temporal embedding.
Two Variants:
1. P2P-PL (Parcel-Level): Models the time series of each anatomical parcel independently using a Longitudinal Transformer, followed by late fusion (scoring and aggregation) across parcels.
2. P2P-SL (Scan-Level): First aggregates parcels into a weighted scan embedding for each visit, then models the progression across visits using a Longitudinal Transformer. This variant is more computationally efficient for long histories.
Efficiency: By decoupling spatial encoding (Stage A) from temporal modeling (Stage B), P2P reduces the computational complexity of global attention. While a standard 4D ViT scales quadratically with the number of tokens and visits ( $O(T^2 N^2)$ ), P2P scales linearly with the number of visits ( $O(T \cdot \sum N_p^2)$ ), making long-term analysis computationally tractable.

3. Key Contributions

High-dimensional Anatomical Grounding: The introduction of an Atlas-guided Parcel Encoder that tokenizes high-resolution 3D scans into atlas-grounded representations. This preserves fine-grained regional specificity and provides richer spatial context than standard global embeddings.
Computationally Efficient Temporal Scaling: A two-stage approach that significantly reduces self-attention overhead. By coupling parcel-specific encoding with an age-aware Longitudinal Transformer, the framework accommodates irregular visit intervals and scales linearly with the number of visits.
Superior Diagnostic and Prognostic Performance: The framework achieves state-of-the-art (SOTA) performance across multiple datasets (ADNI, AIBL, MIRIAD) and demonstrates strong generalization to cross-datasets and other neurodegenerative diseases (Frontotemporal Dementia).

4. Experimental Results

The framework was evaluated on ADNI, AIBL, MIRIAD, and NIFD datasets, as well as a controlled synthetic phantom dataset.

Performance Gains:
- AD vs. CN Classification: P2P outperforms prior single-scan and longitudinal baselines. Leveraging longitudinal data improved balanced accuracy by up to 4% over single-scan baselines.
- MCI-to-AD Conversion (sMCI vs. pMCI): This is a particularly challenging task. P2P achieved significant gains, with balanced accuracy improvements of up to 7.59% and AUC improvements of 6% over single-scan baselines.
- Cross-Dataset Generalization: Models trained on ADNI and tested on AIBL/MIRIAD (via inference or fine-tuning) maintained high performance, indicating the capture of cohort-stable anatomical signals.
Efficiency: Compared to a standard 4D ViT, the parcel grouping strategy reduces self-attention compute costs by approximately 70% (roughly 5x fewer operations) while maintaining or improving performance.
Synthetic Validation: On a synthetic dataset with known ground-truth atrophy locations, P2P successfully localized the abnormal regions. Attention rollouts and gate weights concentrated on the altered parcels, validating the model's ability to identify the spatial origin of predictions.
Risk Stratification: Using Cox proportional hazards models on P2P embeddings, the framework effectively stratified MCI subjects into distinct future conversion-risk groups, with meaningful discrimination observed even at early landmarks.

5. Significance and Claims

The paper claims that P2P addresses a critical gap in neuroimaging by unifying anatomy-aware model design and longitudinal modeling. The authors argue that these are not separate subdomains but two facets of a single region-specific progression phenomenon.

Clinical Relevance: By preserving spatial resolution and handling variable-length histories, P2P enables robust differentiation between stable and progressive MCI, a task where single-time-point snapshots often fail due to overlapping morphological signatures.
Interpretability: The framework offers direct, region-specific interpretability. The parcel scoring mechanism allows clinicians to identify which anatomical regions (e.g., hippocampus, amygdala) contributed most to a prediction, revealing atrophy patterns consistent with established AD literature.
Scalability: The linear scaling with the number of visits allows for the analysis of long-term clinical trajectories without the need for dataset truncation or aggressive downsampling, making the model viable for diverse clinical settings with irregular data collection.

The authors conclude that while P2P is currently limited to T1w MRI and the Harvard-Oxford atlas, its design principles provide a scalable foundation for future integration of multi-modal data (e.g., PET, DTI) and application to other neurodegenerative conditions.