Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Acquisition of interpretable domain information during brain MR image harmonization for content-based image retrieval

Keima Abe, Hayato Muraki, Shuhei Tomoshige, Kenichi Oishi, Hitoshi Iyatomi

TL;DR

This work addresses domain shifts in brain MR images that hinder content-based image retrieval by proposing PL-SE-ADA, a framework that disentangles domain-invariant biological content from domain-specific style. It introduces two encoders, a decoder, and a domain predictor, and reconstructs images via $\boldsymbol{x}' = f_D(\boldsymbol{z_u}) + \alpha f_D(\boldsymbol{z_d})$, enabling separate visualization of disease-relevant features and domain-related variations. Across ADNI2/3 data, PL-SE-ADA achieves competitive reconstruction and domain-harmonization performance while improving disease classification accuracy over SE-ADA, and provides interpretable, image-space representations of both latent factors. The approach supports potential style-transfer applications and enhances trust and usability of CBIR in clinical settings, though it invites further validation across more diverse domains and refinements in reconstruction quality.

Abstract

Medical images like MR scans often show domain shifts across imaging sites due to scanner and protocol differences, which degrade machine learning performance in tasks such as disease classification. Domain harmonization is thus a critical research focus. Recent approaches encode brain images $\boldsymbol{x}$ into a low-dimensional latent space $\boldsymbol{z}$, then disentangle it into $\boldsymbol{z_u}$ (domain-invariant) and $\boldsymbol{z_d}$ (domain-specific), achieving strong results. However, these methods often lack interpretability$-$an essential requirement in medical applications$-$leaving practical issues unresolved. We propose Pseudo-Linear-Style Encoder Adversarial Domain Adaptation (PL-SE-ADA), a general framework for domain harmonization and interpretable representation learning that preserves disease-relevant information in brain MR images. PL-SE-ADA includes two encoders $f_E$ and $f_{SE}$ to extract $\boldsymbol{z_u}$ and $\boldsymbol{z_d}$, a decoder to reconstruct the image $f_D$, and a domain predictor $g_D$. Beyond adversarial training between the encoder and domain predictor, the model learns to reconstruct the input image $\boldsymbol{x}$ by summing reconstructions from $\boldsymbol{z_u}$ and $\boldsymbol{z_d}$, ensuring both harmonization and informativeness. Compared to prior methods, PL-SE-ADA achieves equal or better performance in image reconstruction, disease classification, and domain recognition. It also enables visualization of both domain-independent brain features and domain-specific components, offering high interpretability across the entire framework.

Acquisition of interpretable domain information during brain MR image harmonization for content-based image retrieval

TL;DR

This work addresses domain shifts in brain MR images that hinder content-based image retrieval by proposing PL-SE-ADA, a framework that disentangles domain-invariant biological content from domain-specific style. It introduces two encoders, a decoder, and a domain predictor, and reconstructs images via , enabling separate visualization of disease-relevant features and domain-related variations. Across ADNI2/3 data, PL-SE-ADA achieves competitive reconstruction and domain-harmonization performance while improving disease classification accuracy over SE-ADA, and provides interpretable, image-space representations of both latent factors. The approach supports potential style-transfer applications and enhances trust and usability of CBIR in clinical settings, though it invites further validation across more diverse domains and refinements in reconstruction quality.

Abstract

Medical images like MR scans often show domain shifts across imaging sites due to scanner and protocol differences, which degrade machine learning performance in tasks such as disease classification. Domain harmonization is thus a critical research focus. Recent approaches encode brain images into a low-dimensional latent space , then disentangle it into (domain-invariant) and (domain-specific), achieving strong results. However, these methods often lack interpretabilityan essential requirement in medical applicationsleaving practical issues unresolved. We propose Pseudo-Linear-Style Encoder Adversarial Domain Adaptation (PL-SE-ADA), a general framework for domain harmonization and interpretable representation learning that preserves disease-relevant information in brain MR images. PL-SE-ADA includes two encoders and to extract and , a decoder to reconstruct the image , and a domain predictor . Beyond adversarial training between the encoder and domain predictor, the model learns to reconstruct the input image by summing reconstructions from and , ensuring both harmonization and informativeness. Compared to prior methods, PL-SE-ADA achieves equal or better performance in image reconstruction, disease classification, and domain recognition. It also enables visualization of both domain-independent brain features and domain-specific components, offering high interpretability across the entire framework.
Paper Structure (15 sections, 2 equations, 4 figures, 3 tables)

This paper contains 15 sections, 2 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Pseudo-Linear-SE-ADA (PL-SE-ADA)
  3. Overview of PL-SE-ADA
  4. Training and Behavior of PL-SE-ADA
  5. Experiments
  6. Data Sets and Preprocessing
  7. Model architecture
  8. Evaluation
  9. Impact of hyperparameter $\alpha$
  10. Results and Discussion
  11. Qualitative Results
  12. Quantitative Results
  13. Impact of hyperparameter
  14. Further applications and limitations
  15. Conclusion

Figures (4)

  • Figure 1: Architecture of PL-SE-ADA.
  • Figure 2: Examples of reconstruction with PL-SE-ADA.
  • Figure 3: Visualization of domain-invariant biological representation ($\boldsymbol{z_u}$: left) and domain-dependent representation ($\boldsymbol{z_d}$: right) extracted by PL-SE-ADA and projected using UMAP.
  • Figure 4: Comparison of image reconstruction under varying domain ratios $\alpha$.