Source-linked AI summary
The Domain Shift Problem of Medical Image Segmentation and Vendor-Adaptation by Unet-GAN
Wenjun Yan, Yuanyuan Wang, Shengjia Gu, Lu Huang, Fuhua Yan, Liming Xia, Qian Tao
TL;DR
Domain shifts can make medical image segmentation performance drop unexpectedly, while annotating every new domain is practically impossible. The authors propose Unet-GAN, combining unpaired vendor adaptation with segmentation-specific feature learning, and report improved segmentation across vendors.
Problem
Medical-image segmentation models can lose performance on data from different vendors, centers, or acquisition settings, while annotating every new distribution is practically impossible.
Source: §1.1 · p. 2Method
Unet-GAN combines unpaired GAN-based vendor adaptation with U-Net segmentation, using SSIM image loss and segmentation-specific U-Net feature loss.
Source: §1.1 · p. 2, §2 · p. 4Results
The authors report that Unet-GAN improves segmentation across vendors, with Dice indices reported in Table 1 and examples of degradation and improvement in Figures 4 and 5.
Source: §3.3 · p. 6Takeaways & Limitations
The authors conclude that Unet-GAN can extend a U-Net trained on one vendor to other vendors without new annotation.
Source: §4 · p. 8
Abstract
from arXiv · showhide
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.
1 Introduction
CNNs have advanced image analysis, and U-Net performs strongly in medical segmentation, yet domain variation can undermine deployment. The authors propose Unet-GAN to address this cross-domain problem.
1 Introduction
1.1 The Domain Shift Problem of CNN Segmentation
- Using fast gradient sign perturbations, the authors generated LV-MRI examples in which nearly imperceptible changes misdirected or defeated a trained U-Net.The perturbation is computed from the sign of the loss gradient with respect to the input.Source: §1.1 · p. 2, Figure 1 · p. 2
- Medical-image distribution shifts across centers, vendors, or acquisition parameters can cause a model’s performance to drop unexpectedly on unseen data.The paper contrasts strong performance on same-distribution test data with possible degradation on data from a different distribution.Source: §1.1 · p. 2
- Annotating large datasets for every new distribution is practically impossible, and source-domain augmentation is not an ultimate fix for systematic shifts.Source: §1.1 · p. 2
- Unet-GAN combines an unpaired GAN for vendor adaptation with a U-Net for segmentation, demonstrated on LV segmentation in cine MRI.Source: §1.1 · p. 2
2 Method
Unet-GAN combines a supervised LV segmentation network with unpaired CycleGAN translation, using image- and feature-level losses to adapt images across domains. Its workflow trains the segmenter on annotated source data, learns translation from unannotated source and target data, then translates target images before segmentation.
CycleGAN
- CycleGAN provides unpaired image translation through two generators and two discriminators, with adversarial and cycle-consistency losses.The generators translate between source and target domains, while the discriminators distinguish original from translated images.Source: §2 · p. 3, Figure 2 · p. 3
Unet-GAN: Dedicated to medical image segmentation across domains
- Unet-GAN combines image- and feature-level information in its loss function for medical image adaptation.The image-level term uses SSIM, while the feature-level term matches segmentation-specific encoder features.Source: §2 · p. 4
- The image-level loss replaces CycleGAN’s pixelwise MSE with SSIM, weighting structural similarity more heavily than luminance and contrast.The paper motivates this choice by noting that medical-image adaptation should preserve anatomy even when luminance and contrast vary.Source: §2 · p. 4
- The feature-level loss matches encoder features of original and translated images, encouraging the GAN to preserve features relevant to LV segmentation.The LV-Unet encoder acts as the feature extractor, and the loss is the mean squared error between the two feature representations.Source: §2 · p. 4
Overall workflow
- The workflow trains LV-Unet on annotated source data, trains the generators and discriminators on unannotated data from both domains, then translates target images to the source domain for segmentation.The translated target images are segmented by the LV-Unet trained on the source domain.Source: §2 · pp. 3–5
3 Experiments & Results
The experiments test whether Unet-GAN can restore LV segmentation across vendor shifts, using Philips as the labeled source and Siemens and GE as targets. Translation improved target-domain Dice in both target vendors.
3.1 Data
- The study used cine MRI from 144 subjects across Philips, GE, and Siemens, with Philips as the source and the other vendors as targets.Philips data were split into 35 training and 9 testing subjects for LV-Unet; Siemens and GE were the target domains.Source: §3.1 · p. 5
3.2 Experiments and Performance Evaluation
- Three scenarios compared target-domain segmentation using LV-Unet trained on clean source data, training on source data with random noise, and target-to-source translation before clean-source LV-Unet segmentation, evaluated by Dice overlap index.Two Unet-GANs translated Siemens to Philips (trained with 3008 Philips and 1680 Siemens images) and GE to Philips (3008 Philips and 1320 GE images).Source: §3.2 · pp. 5–6
3.3 Results
- Unet-GAN raised target-domain Dice from 0.474 to 0.805 on Siemens and 0.727 to 0.867 on GE.The comparison includes clean Unet, noisy Unet, and Unet-GAN; the figures show example results before and after translation.Source: Table 1 · p. 6, §3.3 · p. 6, Figure 4 · p. 7
4 Discussion and Conclusion
The discussion frames cross-domain segmentation failure as a consequence of distribution differences, despite good-quality target images. Unet-GAN adapts target images while preserving segmentation-specific features, and the authors conclude it can support cross-vendor generalization without new annotations.
Discussion
- Unet-GAN combines Unet and GAN for cross-domain medical image segmentation, with MRI vendors serving as the demonstrated domains.The paper identifies the clinical-practice vulnerability of CNN segmentation as the motivating problem.Source: §4 · p. 7
- A source-trained CNN can fail on visually good target images when their distribution differs, consistent with the i.i.d. assumption limiting generalization.The discussion says Unet-GAN shifts target distributions toward the source while enforcing similarity of segmentation-specific Unet features.Source: §4 · pp. 7–8
Conclusion
- The authors conclude that Unet-GAN adapts images across vendors so a Unet trained on one vendor can generalize to another without new annotation.They describe extending a trained CNN to multi-center and multi-vendor clinical use as the intended scope.Source: §4 · p. 8