Unsupervised Domain Adaptation via Content Alignment for Hippocampus Segmentation

TL;DR

This work tackles domain shift in cross-domain hippocampus MRI segmentation by separating style and content variations. It introduces an image-space unsupervised domain adaptation framework that first harmonises appearance via $z$-normalisation and histogram standardisation, then learns content-altering, yet anatomically plausible, transformations with a bidirectional deformable image registration backbone guided by a discriminator and ROI-focused segmentation. The method jointly optimises a registration network $\mathcal{R}$, a segmentation network $\mathcal{S}$, and a discriminator $\mathcal{D}$ with a composite loss $\mathcal{L}$ incorporating $\mathcal{L}_{sim}$, $\mathcal{L}_{smooth}$, $\mathcal{L}_{disc}$, and $\mathcal{L}_{seg}$; the transformed data train a downstream segmentation model. Across Morpho-MNIST and three MRI datasets with varying degrees of atrophy, the approach yields consistent improvements over baselines, achieving up to around 24% relative Dice gain in challenging content-shift scenarios, which demonstrates its practical value for robust hippocampus segmentation in diverse populations.

Abstract

Deep learning models for medical image segmentation often struggle when deployed across different datasets due to domain shifts - variations in both image appearance, known as style, and population-dependent anatomical characteristics, referred to as content. This paper presents a novel unsupervised domain adaptation framework that directly addresses domain shifts encountered in cross-domain hippocampus segmentation from MRI, with specific emphasis on content variations. Our approach combines efficient style harmonisation through z-normalisation with a bidirectional deformable image registration (DIR) strategy. The DIR network is jointly trained with segmentation and discriminator networks to guide the registration with respect to a region of interest and generate anatomically plausible transformations that align source images to the target domain. We validate our approach through comprehensive evaluations on both a synthetic dataset using Morpho-MNIST (for controlled validation of core principles) and three MRI hippocampus datasets representing populations with varying degrees of atrophy. Across all experiments, our method outperforms existing baselines. For hippocampus segmentation, when transferring from young, healthy populations to clinical dementia patients, our framework achieves up to 15% relative improvement in Dice score compared to standard augmentation methods, with the largest gains observed in scenarios with substantial content shift. These results highlight the efficacy of our approach for accurate hippocampus segmentation across diverse populations.

Figures (5)

• Figure 1: Comparison of hippocampal volumes between three populations. DRYAD: young and healthy; HarP: dementia research population; OBHC: dementia clinic population. Demonstrating the shrinking of the hippocampi across populations due to age and disease status/severity.
• Figure 2: Top: Overview of the proposed framework, showing the sequence of steps for input source ($\bm{\mathcal{X}}_s$) and target ($\bm{\mathcal{X}}_t$) images to output style and content aligned images that are used to train a downstream segmentation network. Bottom: Proposed model architecture for learning content alignment, formed of registration ($\mathcal{R}$), segmentation ($\mathcal{S}$) and discriminator ($\mathcal{D}$) networks for joint training.
• Figure 3: Example of simulated Morpho-MNIST thick and thin pairs.
• Figure 4: Examples of content alignment for Morpho-MNIST thick$\rightarrow$thin (top row), DRYAD$\rightarrow$HarP (middle row), DRYAD$\rightarrow$OBHC (bottom row).
• Figure 5: Predicted segmentation masks and corresponding dice scores comparing our method to image-space baselines tested on a target sample from HarP (orange) and OBHC (green). GT: Ground Truth; OoD: Out-of-Domain; Affine: Affine augmentations; Deform: Deformable augmentation.