[Paper Review] Self-Training Based Unsupervised Cross-Modality Domain Adaptation for Vestibular Schwannoma and Cochlea Segmentation

Hyungseob Shin, Hyeongyu Kim+4 more|arXiv (Cornell University)|Sep 22, 2021

Ultrasonics and Acoustic Wave Propagation7 references4 citations

TL;DR

This paper proposes a self-training-based unsupervised cross-modality domain adaptation framework for segmenting vestibular schwannoma and cochlea in high-resolution T2-weighted MRI scans using only T1-weighted MRI annotations. By leveraging contrast conversion, pseudo-labeling, and joint training on real and generated T2 scans, the method achieves state-of-the-art performance on the CrossMoDA2021 challenge, with a mean Dice score of 0.8570 for VS and 0.8446 for cochlea.

ABSTRACT

With the advances of deep learning, many medical image segmentation studies achieve human-level performance when in fully supervised condition. However, it is extremely expensive to acquire annotation on every data in medical fields, especially on magnetic resonance images (MRI) that comprise many different contrasts. Unsupervised methods can alleviate this problem; however, the performance drop is inevitable compared to fully supervised methods. In this work, we propose a self-training based unsupervised-learning framework that performs automatic segmentation of Vestibular Schwannoma (VS) and cochlea on high-resolution T2 scans. Our method consists of 4 main stages: 1) VS-preserving contrast conversion from contrast-enhanced T1 scan to high-resolution T2 scan, 2) training segmentation on generated T2 scans with annotations on T1 scans, and 3) Inferring pseudo-labels on non-annotated real T2 scans, and 4) boosting the generalizability of VS and cochlea segmentation by training with combined data (i.e., real T2 scans with pseudo-labels and generated T2 scans with true annotations). Our method showed mean Dice score and Average Symmetric Surface Distance (ASSD) of 0.8570 (0.0705) and 0.4970 (0.3391) for VS, 0.8446 (0.0211) and 0.1513 (0.0314) for Cochlea on CrossMoDA2021 challenge validation phase leaderboard, outperforming most other approaches.

Motivation & Objective

• To address the high cost of annotating multi-contrast MRI data in medical imaging.
• To enable accurate segmentation of vestibular schwannoma and cochlea in T2-weighted MRI without requiring T2 annotations.
• To reduce domain shift between T1 and T2 MRI modalities using unsupervised domain adaptation.
• To improve generalization by combining real T2 scans with pseudo-labeled data and generated T2 scans with true annotations.
• To achieve high segmentation performance in a low-resource annotation setting using self-training.

Proposed method

• A contrast-preserving image translation network converts contrast-enhanced T1-weighted MRI to synthetic high-resolution T2-weighted MRI.
• A segmentation network is first trained on the generated T2 scans using T1 scan annotations.
• Pseudo-labels are predicted on real, non-annotated T2 scans using the trained segmentation model.
• The segmentation model is fine-tuned on a combined dataset of real T2 scans with pseudo-labels and generated T2 scans with ground-truth labels.
• Self-training is iteratively applied to improve robustness and reduce domain shift between T1 and T2 modalities.
• The framework leverages cycle-consistency and adversarial loss in the image translation to preserve anatomical structures.

Experimental results

Research questions

• RQ1 Can self-training effectively reduce domain shift between T1 and T2 MRI for medical image segmentation?
• RQ2 How well can a model trained on T1 annotations generalize to T2 scans without any T2 annotations?
• RQ3 Can pseudo-labeling on real T2 scans improve segmentation performance when combined with synthetic T2 data?
• RQ4 Does joint training on real and generated T2 scans enhance model generalization and robustness?
• RQ5 What is the performance of this approach compared to existing unsupervised domain adaptation methods on the CrossMoDA2021 benchmark?

Key findings

• The method achieved a mean Dice score of 0.8570 (±0.0705) for vestibular schwannoma segmentation on the CrossMoDA2021 validation set.
• The method achieved a mean Dice score of 0.8446 (±0.0211) for cochlea segmentation, outperforming most competing approaches.
• The Average Symmetric Surface Distance (ASSD) for vestibular schwannoma was 0.4970 (±0.3391), indicating high segmentation accuracy.
• The ASSD for cochlea was 0.1513 (±0.0314), reflecting precise boundary localization.
• The model outperformed most other entries in the CrossMoDA2021 challenge, demonstrating state-of-the-art performance.
• The combination of synthetic data and pseudo-labeling significantly improved generalization in the absence of T2 annotations.