PL-SE-ADA: Pseudo-Linear Adversarial Domain Adaptation

• The paper introduces a novel framework using dual encoders to disentangle domain-invariant and domain-specific features, achieving improved interpretability.
• It employs a three-stage adversarial training protocol that promotes domain unlearning while preserving critical clinical features through reconstruction metrics.
• The pseudo-linear additive reconstruction enables explicit manipulation of domain biases, facilitating reliable harmonization and downstream analysis in medical imaging.

Pseudo-Linear-Style Encoder Adversarial Domain Adaptation (PL-SE-ADA) is a domain harmonization and representation learning framework that achieves disentanglement and interpretable decomposition of domain-invariant and domain-specific components in data, enabling robust transfer and downstream analysis under domain shift. In recent developments, the approach has been shown to extend the capabilities of adversarial adaptation by pairing architectural innovations—specifically dual encoders and a pseudo-linear additive reconstruction strategy—with adversarial objectives, achieving both performance and interpretability in applications such as brain MR image harmonization for machine learning and content-based image retrieval (Abe et al., 16 Oct 2025).

1. Architectural Foundations and Model Components

PL-SE-ADA comprises two specialized encoders, a decoder, and a domain predictor:

• Domain-Invariant Encoder ($f_E$): Extracts latent vector $\boldsymbol{z_u}$, capturing biological and disease-relevant information which persists across domain shifts (e.g., anatomical structures in MRI regardless of scanner differences).
• Style Encoder ($f_{SE}$): Produces a low-dimensional site/scanner-specific representation $\boldsymbol{z_d^\prime}$, expanded via affine transformation to a full domain-specific latent code $\boldsymbol{z_d}$, which captures acquisition artifacts.
• Decoder ($f_D$): Generates separate reconstructions from $\boldsymbol{z_u}$ and $\boldsymbol{z_d}$, uniquely enabling partitioned visualizations and arithmetic manipulation—these are recombined additively for final output.
• Domain Predictor ($g_D$): A classifier (MLP) trained adversarially against $f_E$ to force the decorrelation of domain information from $\boldsymbol{z_u}$, supporting harmonization.

This modular design, with explicit separation of domain-relevant and domain-nuisance signals, distinguishes PL-SE-ADA from previous adversarial adaptation schemes which typically intertwine the latent code prior to decoding.

2. Latent Representation Disentanglement

The input image $\boldsymbol{x}$ is mapped to a latent space $\boldsymbol{z}$ split as:

• $\boldsymbol{z_u}$: domain-invariant, containing content necessary for disease classification and anatomical reasoning.
• $\boldsymbol{z_d}$: domain-specific, capturing scanner-dependent variation.

The encoders process $\boldsymbol{x}$ independently, formalized as: $$\boldsymbol{z_u} = f_E(\boldsymbol{x}),\quad \boldsymbol{z_d} = \mathrm{Affine}(f_{SE}(\boldsymbol{x}))$$ where $\mathrm{Affine}(\cdot)$ expands $\boldsymbol{z_d^\prime}$ to match the target latent dimensionality.

This separation is enforced both architecturally and via targeted learning objectives; only $\boldsymbol{z_u}$ is adversarially regularized against the domain, while $\boldsymbol{z_d}$ is allowed to retain domain-specific information, confirmed via clustering and visualization methods (e.g., UMAP projections).

3. Adversarial Domain Unlearning and Training Protocol

PL-SE-ADA adopts a three-stage adversarial training procedure:

1. Reconstruction Training: Encoders ($f_E$, $f_{SE}$) and decoder ($f_D$) are optimized to minimize the reconstruction error:

$$\mathcal{L}_{\text{recon}} = \| \boldsymbol{x} - [f_D(\boldsymbol{z_u}) + \alpha f_D(\boldsymbol{z_d})] \|^2$$

with hyperparameter $\alpha$ controlling the weight of the domain-specific component.

1. Domain Predictor Training: Keeping encoders and decoder fixed, $g_D$ is trained to classify imaging domain from features.
2. Adversarial Encoder Update: With $g_D$ fixed, $f_E$ is adversarially trained to make domain labels from $\boldsymbol{z_u}$ unpredictable, typically by targeting uniform probability over domain classes. This is operationalized via gradient reversal or uniform labeling strategies.

This process purges domain cues from $\boldsymbol{z_u}$, driving harmonization while preserving biological content necessary for clinically relevant analysis.

4. Pseudo-Linear Reconstruction and Interpretability

PL-SE-ADA introduces a pseudo-linear additive reconstruction strategy: $$\boldsymbol{x'} = f_D(\boldsymbol{z_u}) + \alpha f_D(\boldsymbol{z_d}) = \boldsymbol{x_u} + \alpha \boldsymbol{x_d}$$

• $\boldsymbol{x_u}$: the domain-invariant reconstruction, ideal for downstream clinical analysis.
• $\boldsymbol{x_d}$: the domain-specific artifact, visually subtle (e.g., "white haze" in MRI), quantifying scanner-dependent signal.

This innovation enables explicit visual inspection and arithmetic manipulation; for example, varying $\alpha$ modulates the contribution of domain bias, allowing the assessment of harmonization efficacy. Unlike prior approaches (e.g., SE-ADA, which decoded $z_u + z_d$ jointly (Abe et al., 16 Oct 2025)), the separate decodings provide two distinct, interpretable images.

5. Empirical Performance and Evaluation Metrics

PL-SE-ADA is assessed via:

• Image Reconstruction Quality: RMSE and SSIM between $\boldsymbol{x}$ and $\boldsymbol{x'}$ gauge fidelity.
• Disease Classification: Macro F1-score on disease/health status predicted from $\boldsymbol{z_u}$ reflects retention of clinical content.
• Domain Recognition: Macro F1-score of $g_D$ operating on $\boldsymbol{z_u}$ targets 0.5 (chance), indicating successful domain unlearning. $\boldsymbol{z_d}$ is expected to cluster according to domain, confirmed by unsupervised projections.
• Visualization: Separate reconstructions $\boldsymbol{x_u}$ and $\boldsymbol{x_d}$ support clinical transparency, with anatomical features (e.g., cortical folds) preserved in $\boldsymbol{x_u}$ and domain artifacts isolated in $\boldsymbol{x_d}$.

PL-SE-ADA matches or surpasses prior harmonization methods in these metrics, achieving improved disease classification and interpretability without sacrificing reconstruction quality.

6. Comparative Context and Theoretical Significance

Within the ecosystem of adversarial domain adaptation:

• PL-SE-ADA shares architectural motifs with prior frameworks including ADDA (Tzeng et al., 2017), disentangled adaptation with feature exchange (Vu et al., 2017), and the use of adversarial unlearning (Abe et al., 16 Oct 2025).
• The pseudo-linear reconstruction distinguishes PL-SE-ADA, enhancing interpretability without entangling latent signals.
• Its adversarial protocols align with domain-invariant approaches in unsupervised and semi-supervised settings, but offer additional transparency critical for medical and multi-center studies.
• Contextualizing with recent trends in relaxed alignment (Wu et al., 2019), PL-SE-ADA addresses the need for principled handling of domain bias without sacrificing the transferability of intrinsic content.

7. Applications and Implications

PL-SE-ADA is particularly suited for medical image harmonization (notably brain MRI), content-based image retrieval, and any domain adaptation problem where interpretability of latent factors is non-negotiable. The decomposition supports:

• Reliable disease classification in the presence of acquisition bias.
• Visual assessment of harmonization, facilitating trust in downstream analysis.
• The modification of domain-specific contributions (via $\alpha$) for diagnostic or algorithmic calibration.

A plausible implication is that similar pseudo-linear strategies could improve interpretability in other domains by supporting disentangled additive decompositions and targeted adversarial unlearning.

---

PL-SE-ADA thus advances the field of adversarial domain adaptation by combining disentangled latent modeling, interpretable pseudo-linear reconstruction, and robust adversarial harmonization in a unified framework suitable for high-stakes, bias-sensitive domains such as medical image analysis (Abe et al., 16 Oct 2025).