Unsupervised Cross-modality Domain Adaptation for Segmenting Vestibular Schwannoma and Cochlea with Data Augmentation and Model Ensemble

Hao Li, Dewei Hu, Qibang Zhu, Kathleen E. Larson, Huahong Zhang, Ipek Oguz

Introduction

Deep learning has revolutionized medical image segmentation, yet its reliance on large, manually labeled datasets remains a significant bottleneck. In clinical practice, labels are often available for one modality but not another. For instance, Vestibular Schwannoma (VS) and the cochlea are typically labeled in contrast-enhanced T1-weighted (ceT1) MRI, while high-resolution T2-weighted scans are frequently used for longitudinal monitoring. Applying models trained on ceT1 directly to T2 scans often fails due to "domain shift"—variations in contrast, resolution, and noise. This paper presents an unsupervised domain adaptation (UDA) framework to bridge this gap, enabling accurate segmentation of T2 scans without requiring target-domain labels.

Methodology

The proposed framework follows a two-stage pipeline: Synthesis (Syn) and Segmentation (Seg).

Synthesis Stage

To address the lack of labeled T2 data, the researchers utilized a CycleGAN for unpaired image-to-image translation. This model learns to transform labeled ceT1 images into "pseudo-T2" images. These synthetic images mimic the visual characteristics of real T2 scans while preserving the precise anatomical labels from the original ceT1 data.

Segmentation Stage

The segmentation phase employs an ensemble approach to ensure robustness against synthetic artifacts and anatomical variability. The ensemble consists of four deep learning models:

  • Two 2.5D CNNs: These models utilize 2D convolutions in early layers and 3D convolutions later, effectively managing the anisotropic resolution common in clinical MRI.
  • Two 3D CNNs: One model incorporates an attention module to focus on small anatomical structures (like the cochlea), while the other serves as a standard 3D baseline.
Overview of the Syn-Seg pipeline, illustrating the transition from 3D ceT1 MRI to synthetic T2 images and the subsequent ensemble-based segmentation.

Data Augmentation and Preprocessing

To bridge the "reality gap" between synthetic and real T2 images, heavy online data augmentation was applied, including elastic deformations, gamma correction, and the addition of Gaussian noise and blur. A critical preprocessing step involved aligning all images to the MNI brain template and cropping them to a specific Region of Interest (ROI) centered on the internal auditory canal. This focused the networks on the small target structures, significantly improving performance.

Experimental Results

The framework was evaluated using the cross-MoDA challenge dataset, comprising 105 labeled ceT1 and 105 unlabeled T2 scans for training, with 32 T2 scans for validation.

Quantitative Performance

The results demonstrate that the ensemble approach, combined with ROI cropping, significantly outperforms baseline methods. ROI cropping alone improved the VS Dice score from 0.408 to 0.662, while the final ensemble achieved state-of-the-art results for unsupervised adaptation.

Method VS Dice Score Cochlea Dice Score
Baseline (No cropping) 0.408 0.405
Crop + Augmentation 0.740 0.731
Proposed (Ensemble) 0.794 0.741

Qualitative Analysis

Visual inspection of the segmentations confirms that the model accurately captures the boundaries of both the VS (red) and the cochlea (green), even in cases with varying tumor sizes and locations.

Qualitative segmentation results across different test cases, showing the model's ability to accurately delineate the Vestibular Schwannoma (red) and cochlea (green).

Discussion and Conclusion

The success of this framework highlights two key insights: the necessity of focused ROI cropping for small anatomical structures and the robustness provided by model ensembles. While 2.5D models handle resolution issues well, 3D models provide better global context; their fusion mitigates individual weaknesses.

Future work will address the limitations of 2D CycleGANs, which can introduce inter-slice intensity inconsistencies. Additionally, the researchers aim to implement self-training strategies, using the current model to generate pseudo-labels on real T2 data for iterative refinement. This approach demonstrates that unsupervised domain adaptation is a viable and promising path for clinical applications where labeled data is scarce.

Summary updated

BibTeX
@inproceedings{li2021unsupervised,
  author = {Hao Li and Dewei Hu and Qibang Zhu and Kathleen E. Larson and Huahong Zhang and Ipek Oguz},
  title = {{Unsupervised Cross-modality Domain Adaptation for Segmenting Vestibular Schwannoma and Cochlea with Data Augmentation and Model Ensemble}},
  booktitle = {BrainLes @ MICCAI},
  year = {2021},
  url = {https://openreview.net/forum?id=VRs8dV2ui7},
}
Figure