🤖 machine learning

CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation

The paper proposes CASL-VAE, a deep contrastive latent variable model that learns structured continuous and discrete latent factors from unpaired data to enable semi-supervised clustering, paired-sample generation, and the analysis of population heterogeneity in clinical contexts like Alzheimer's disease.

Original authors: Sai Spandana Chintapalli, Pratik Chaudhari, Christos Davatzikos

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sai Spandana Chintapalli, Pratik Chaudhari, Christos Davatzikos

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 you are a detective trying to figure out why a group of people (let's call them the "Target Team") looks different from a group of healthy, normal people (the "Reference Team"). Usually, you'd want to see the same person before and after they got sick to spot the changes. But in the real world, you rarely have those perfect "before-and-after" photos. You just have a pile of healthy faces and a pile of sick faces, and you don't know which sick person matches which healthy person.

This is where the paper introduces CASL-VAE. Think of it as a super-smart, magical sorting machine that can learn from these unpaired piles of data without getting confused.

The Problem: The Messy Entanglement

Usually, when scientists try to compare these two groups, the results get messy.

The Noise: Healthy people naturally have differences (some are tall, some are short, some have different brain shapes just because of their age or scanner). The machine often mistakes these normal differences for disease signs.
The Variety: The "Target Team" isn't just one big blob of sickness. They might have different types of the same problem (like different flavors of ice cream, all labeled "sick"). Old methods struggle to separate these flavors.
The Missing Link: Without paired data (where you know exactly who matches who), standard tools get tangled up and can't tell what is "shared" between the groups and what is "unique" to the sick group.

The Solution: The Two-Drawer Filing System

The authors built CASL-VAE, which acts like a smart filing system with two special drawers to organize the information:

The "Common" Drawer: This holds everything that is shared between healthy and sick people. Think of this as the "baseline" stuff—like the fact that everyone has a brain, or that everyone's brain shrinks a little bit as they age. The machine learns to put all these shared traits here so they don't get in the way.
The "Salient" (Important) Drawer: This is where the magic happens. This drawer is split into two parts:
- The "Category" Slot: This sorts the sick people into distinct subgroups (like "Flavor A," "Flavor B," "Flavor C"). The machine discovers these groups on its own, without anyone telling it what they are.
- The "Fine-Tune" Slot: Even within "Flavor A," people aren't identical. This slot captures the small, continuous differences between individuals in that same group.

How It Works Without "Before-and-After" Photos

The paper suggests a clever trick. Since the machine can't see the direct link between a specific healthy person and a specific sick person, it forces the "Common Drawer" to look the same whether it's looking at a healthy person or a sick person. It's like saying, "No matter which team you are looking at, the 'base brain' structure must be consistent."

By strictly separating the "shared" stuff from the "unique sick" stuff, the machine can learn the patterns of the disease even without matched pairs. It essentially says, "Okay, I know what a normal brain looks like. Now, I'll look at this sick brain, strip away the normal parts, and see what's left over. That leftover stuff is the disease."

What the Paper Actually Found (and Didn't Find)

The authors tested this idea in two ways:

1. The Simulation (The "Fake" Test)
They created a fake dataset using real brain scans from healthy people and then artificially "sickened" them by shrinking specific brain regions by 10–30%. They created three fake disease subtypes.

The Result: In these simulations, CASL-VAE did a great job finding the fake subtypes. It scored an Adjusted Rand Index (ARI) of 0.620 (on a scale where 1 is perfect) in a typical scenario, which was better than most other methods. In a harder, "mild" scenario (only 10–20% shrinkage), it scored 0.468, beating all other models significantly.
The Visuals: The paper shows that the patterns the machine "learned" looked almost exactly like the fake patterns they programmed in. It also showed that it could generate a "paired" healthy version of a sick brain, and the similarity score was around 0.58, which was higher than the other models.

2. The Real World Test (Alzheimer's)
They then applied this to real data from the ADNI dataset, comparing Cognitively Normal (CN) people against those with Mild Cognitive Impairment (MCI) or Alzheimer's Disease (AD).

The Result: The paper suggests that the machine successfully found hidden subtypes within the Alzheimer's group that made biological sense. However, the paper does not claim to have "solved" Alzheimer's or proven a new cure. It simply suggests that this method can reveal hidden structures in the data that other methods might miss.

What the Paper Explicitly Rules Out

The authors are very clear about what their method is not:

It is not a method that requires paired data. If you don't have matched healthy/sick pairs, this is the tool to use.
It is not just a simple clustering tool (like K-means). Simple clustering often gets confused by the "shared" noise (like age differences), but CASL-VAE specifically filters that out first.
It does not assume the disease is just one single thing. It argues against methods that treat all sick people as a single group, showing instead that there are likely multiple distinct subtypes.

The Bottom Line

The paper suggests that by using this structured, two-drawer approach, we can finally untangle the messy differences between healthy and sick populations, even when we don't have perfect matching data. In their simulations, it worked better than existing tools at finding hidden subtypes and generating paired examples. In the real world, it hints at a way to see the different "flavors" of Alzheimer's disease more clearly, but the authors present this as a promising step forward in understanding, not a final solution.

Technical Summary: CASL-VAE

Problem Statement
Many scientific and clinical challenges require quantifying variability in a target population relative to a reference population (e.g., diseased vs. healthy). However, existing methods face three primary obstacles in this setting:

Entanglement: Target-specific variation is often confounded with variation shared across domains, obscuring subtle effects.
Heterogeneity: Target populations frequently exhibit multi-modal heterogeneity (e.g., distinct disease subtypes) that standard models fail to disentangle.
Lack of Paired Data: Direct supervision via paired observations (where a specific target sample corresponds to a specific reference sample) is rarely available, preventing direct learning of cross-domain relationships.

Consequently, current generative models often entangle shared and domain-specific factors, while standard contrastive analysis methods lack the latent structure necessary to capture multi-modal target heterogeneity or support paired-sample generation.

Methodology: CASL-VAE
The authors propose CASL-VAE (Contrastive Analysis with Structured Latents via Variational Autoencoder), a deep generative framework designed to learn structured, disentangled latent representations from unpaired reference–target data.

Latent Factorization: The model decomposes the latent space into two distinct components:
1. Continuous Common Latent ( $z$ ): A domain-invariant variable capturing variation shared across both reference and target populations.
2. Hierarchical Salient Latent ( $s$ ): A target-specific variable modeling heterogeneous variation. This is further factorized into:
  - A discrete component ( $s_{cat}$ ) representing latent subtypes (e.g., disease subtypes).
  - A continuous component ( $s_{cont}$ ) conditioned on the discrete component, capturing fine-grained variation within each subtype.
Generative Process: The model assumes that target observations ( $x$ ) are generated from both $z$ and $s$ , while reference observations ( $y$ ) are generated solely from $z$ (with salient latents fixed to uninformative preset values). This additive structure allows for the explicit separation of common and salient contributions in the input space.
Optimization from Unpaired Data: A core theoretical contribution is the derivation of a variational lower bound for the joint log-likelihood using only marginal (unpaired) observations. By imposing a posterior decoupling constraint ( $q(z|x) = q(z|y)$ ), the common latent becomes domain-invariant. This allows the model to optimize an objective equivalent to maximizing a relaxed joint likelihood without requiring paired training samples.
Regularization: To ensure the salient latent captures only target-specific variation and the common latent remains domain-invariant, the authors employ four regularization terms:
1. Maximum Mean Discrepancy (MMD): Aligns the distributions of the common latent across domains.
2. Sample-level Latent Consistency: Enforces stability of the common latent under cross-domain transformations.
3. Semi-supervised Binary Cross-Entropy: Uses domain labels (reference vs. target) as weak supervision to force reference samples into a specific discrete category, leaving other categories for unsupervised target subtype discovery.
4. Reference Regularization: Suppresses information in the continuous salient latent for reference samples, preventing it from encoding reference-domain variation.

Key Contributions

Framework: Introduction of CASL-VAE, a contrastive generative framework that learns disentangled shared and target-specific representations from unpaired data using structured latent variables.
Hierarchical Decomposition: Proposal of a hierarchical salient latent structure that enables unsupervised discovery of target subtypes while preserving continuous within-subtype variation.
Downstream Capabilities: Demonstration that this structure enables interpretable downstream tasks, specifically paired-sample generation (synthesizing a reference sample for a given target, or vice versa) and cross-domain comparison, despite the absence of paired training data.

Experimental Results
The authors validated CASL-VAE on neuroimaging data (brain volumes from the iSTAGING consortium and ADNI dataset).

Semi-synthetic Experiments: Using simulated disease patterns with known ground-truth subtypes and atrophy rates (10–30%):
- Subtype Discovery: CASL-VAE achieved superior Adjusted Rand Index (ARI) scores compared to baselines (including GMM, K-means, cVAE, SepVAE, and Smile-GAN), particularly in "mild" atrophy settings (10–20%) where effect sizes were weaker. It successfully distinguished target subtypes from the reference domain in joint clustering.
- Interpretability: The model recovered prototypical target patterns that showed near-perfect correspondence with simulated ground-truth masks.
- Paired Generation: In paired-sample generation tasks, CASL-VAE achieved higher cosine similarity between generated and ground-truth paired reference samples compared to cVAE and SepVAE.
Real Data Application: Applied to an Alzheimer's Disease (AD) cohort (CN as reference, MCI/AD as target), the model demonstrated the ability to reveal biologically plausible heterogeneity, successfully identifying distinct subtypes within the AD population without prior subtype labels.

Significance and Claims
The paper claims that CASL-VAE provides a principled probabilistic basis for performing cross-domain analysis and paired-sample generation using only unpaired data. By integrating contrastive analysis with structured hierarchical latent variables, the method addresses the limitations of existing approaches that either entangle factors or fail to model multi-modal target heterogeneity. The authors emphasize that their approach enables unsupervised target subtype discovery while maintaining interpretability, distinguishing it from prior contrastive and clustering methods that rely on post-hoc analyses or pre-existing subtype labels. The framework is presented as broadly applicable to any setting requiring unpaired reference–target comparisons and the modeling of heterogeneous target variation.