Robust White Matter Hyperintensity Segmentation On Unseen Domain

Xingchen Zhao, Anthony Sicilia, Davneet S. Minhas, Erin O'Connor, Howard J. Aizenstein, William E. Klunk, Dana L. Tudorascu, Seong Jae Hwang

Introduction

In medical imaging, a significant challenge is the "domain shift" that occurs when a machine learning model trained on data from one hospital is deployed at another. Variations in MRI scanner hardware (e.g., 1.5T vs. 3T field strengths), imaging protocols, and manufacturer-specific software create systematic differences in image intensity and contrast. Standard models, often referred to as "DeepAll," frequently fail in these new environments because they overfit to the specific characteristics of their training data.

While Domain Adaptation techniques can mitigate this by fine-tuning models on data from the target site, this requires prior access to such data. This paper addresses the more challenging problem of Domain Generalization (DG): developing a model robust enough to perform accurately on a completely "unseen" domain out of the box, without any prior knowledge of the target scanner's characteristics.

Methods: The MixDANN Framework

The authors propose MixDANN, a hybrid framework built upon the standard U-Net architecture. It integrates two complementary techniques to achieve domain invariance:

  1. Domain Adversarial Neural Network (DANN): This component incorporates a domain discriminator that attempts to identify the source scanner of a given image. The feature extractor is trained to "fool" this discriminator, thereby learning features that are invariant to site-specific variations.
  2. Mixup Data Augmentation: This technique involves training the model on convex combinations of images and labels from different domains (e.g., blending images from Scanner A and Scanner B).

The core innovation lies in the synergy between these methods. Mathematically, both DANN and Mixup work to minimize the $\mathcal{H}$-divergence between distributions. Mixup effectively "fills the gaps" between known training domains, enabling the DANN component to learn a smoother and more generalized representation of the data.

The MixDANN architecture integrates a domain discriminator and mixup augmentation into a standard U-Net.

Experimental Setup

The study utilized MRI data (FLAIR and T1 sequences) from two primary sources:

  • MICCAI WMH Segmentation Challenge Dataset: Includes data from Amsterdam (A), Singapore (S), and Utrecht (U).
  • In-house Dataset: Data from a single site in Pittsburgh (P).

Evaluation followed a "leave-one-site-out" protocol. For instance, the model might be trained on Amsterdam and Singapore data and then tested on the entirely unseen Utrecht dataset to assess its generalization capabilities.

Results

The performance was primarily evaluated using the Dice Similarity Coefficient (DSC). MixDANN demonstrated substantial improvements over baseline models on unseen data.

  • Generalization Performance: In the "A+S → U" test case, the baseline U-Net achieved a DSC of only 0.183. MixDANN significantly improved this to 0.694.
  • Volume Accuracy: For clinical applications, the Absolute Volume Difference (AVD) is a critical metric. MixDANN reduced the volume error from 150.24 (baseline) to 20.89 on the challenge dataset.
  • Leaderboard Ranking: Despite using a single model rather than a complex ensemble, MixDANN reached 6th place on the global WMH Challenge leaderboard.
  • Qualitative Comparison: Visual inspections confirm that MixDANN produces segmentations that are much closer to the ground truth compared to baseline and individual component methods.
Qualitative comparison of segmentation results across different methods on unseen domain data.

Discussion and Limitations

While MixDANN shows strong performance in intensity-based segmentation, the authors noted a drop in "counting" metrics (Recall and F1) when testing on the in-house Pittsburgh dataset. This discrepancy is attributed to differing annotation standards across institutions—what one radiologist identifies as a lesion may differ from another's definition.

Future research will focus on making models robust not only to scanner variations but also to variations in human labeling styles, ensuring consistent "ground truth" remains consistent across the medical community.

Summary updated

BibTeX
@inproceedings{zhao2021robust,
  author = {Xingchen Zhao and Anthony Sicilia and Davneet S. Minhas and Erin O'Connor and Howard J. Aizenstein and William E. Klunk and Dana L. Tudorascu and Seong Jae Hwang},
  title = {{Robust White Matter Hyperintensity Segmentation On Unseen Domain}},
  booktitle = {ISBI},
  year = {2021},
  doi = {10.1109/isbi48211.2021.9434034},
}

Loading related content...

Figure