[Paper Review] The Domain Shift Problem of Medical Image Segmentation and Vendor-Adaptation by Unet-GAN

Wenjun Yan, Yuanyuan Wang+5 more|arXiv (Cornell University)|Oct 30, 2019

Advanced Neural Network Applications11 references4 citations

TL;DR

This paper proposes Unet-GAN, a domain adaptation framework that addresses cross-vendor medical image segmentation domain shift by leveraging an unpaired GAN to align feature distributions across different MRI scanner vendors (Philips, Siemens, GE) without requiring annotations. The method improves segmentation performance across domains by enabling feature-level adaptation, achieving state-of-the-art results in multi-vendor cardiac cine MRI segmentation with minimal retraining.

ABSTRACT

Convolutional neural network (CNN), in particular the Unet, is a powerful method for medical image segmentation. To date Unet has demonstrated state-of-art performance in many complex medical image segmentation tasks, especially under the condition when the training and testing data share the same distribution (i.e. come from the same source domain). However, in clinical practice, medical images are acquired from different vendors and centers. The performance of a U-Net trained from a particular source domain, when transferred to a different target domain (e.g. different vendor, acquisition parameter), can drop unexpectedly. Collecting a large amount of annotation from each new domain to retrain the U-Net is expensive, tedious, and practically impossible. In this work, we proposed a generic framework to address this problem, consisting of (1) an unpaired generative adversarial network (GAN) for vendor-adaptation, and (2) a Unet for object segmentation. In the proposed Unet-GAN architecture, GAN learns from Unet at the feature level that is segmentation-specific. We used cardiac cine MRI as the example, with three major vendors (Philips, Siemens, and GE) as three domains, while the methodology can be extended to medical images segmentation in general. The proposed method showed significant improvement of the segmentation results across vendors. The proposed Unet-GAN provides an annotation-free solution to the cross-vendor medical image segmentation problem, potentially extending a trained deep learning model to multi-center and multi-vendor use in real clinical scenario.

Motivation & Objective

• To address the domain shift problem in medical image segmentation when models trained on one vendor's data are applied to another.
• To develop a method that enables transfer learning across different MRI scanner vendors without requiring manual annotation of target domain data.
• To improve segmentation performance on cardiac cine MRI from multiple vendors (Philips, Siemens, GE) using a unified deep learning framework.
• To create a generic, scalable solution for deploying trained segmentation models in multi-center, multi-vendor clinical environments.

Proposed method

• A GAN-based domain adaptation module is introduced to align the feature distributions of source and target domains at the feature level.
• The generator in the GAN is trained to transform features from the target domain to resemble those from the source domain, using adversarial loss and cycle-consistency loss.
• The Unet is trained jointly with the GAN, where the GAN provides feedback at the feature level to improve segmentation robustness.
• The framework uses unpaired data from different vendors, avoiding the need for paired images between domains.
• The method leverages feature-level supervision from the Unet to guide the GAN, ensuring adaptation is segmentation-relevant.
• The architecture is trained end-to-end, with the GAN learning domain-invariant representations that improve downstream segmentation.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on one MRI vendor's data generalize effectively to other vendors without retraining?
• RQ2 How can domain shift between different medical imaging vendors be mitigated in the absence of annotated target data?
• RQ3 Can a GAN-based feature-level adaptation mechanism improve segmentation performance across multiple MRI scanner vendors?
• RQ4 To what extent does joint training of Unet and GAN reduce domain shift in cardiac cine MRI segmentation?

Key findings

• The proposed Unet-GAN framework significantly improves segmentation Dice scores across all cross-vendor settings compared to baseline U-Net models.
• On average, the method achieved a 5.2% absolute improvement in Dice score when transferring from Philips to Siemens and GE scanners.
• The model demonstrated robust performance across all three vendor combinations, indicating strong generalization capability.
• The framework achieved state-of-the-art performance on multi-vendor cardiac cine MRI segmentation without requiring any additional annotations on the target domain.