Source-linked AI summary
Semantic-Aware Generative Adversarial Nets for Unsupervised Domain Adaptation in Chest X-ray Segmentation
Cheng Chen, Qi Dou, Hao Chen, Pheng-Ann Heng
TL;DR
Domain shift reduces the generalization of medical-image segmentation models, while supervised adaptation requires labeled target data. SeUDA transforms target images into source-like appearances using semantic-aware GANs that preserve structure, and its unsupervised lung-segmentation performance was highly competitive with supervised transfer learning.
Problem
Domain shift degrades DNN segmentation on new datasets, while supervised transfer learning relies on expensive or infeasible labeled target data.
Method
SeUDA uses GAN-based target-to-source image transformation with cycle-consistency and nested semantic-aware adversarial learning, while keeping the source segmenter fixed and using no target labels.
Results
SeUDA exceeded the UDA baseline and was highly competitive with supervised transfer learning for lung segmentation across two public chest X-ray datasets.
Takeaways & Limitations
A source-domain segmentation DNN can be directly generalized to transformed target images without training a new model for each target dataset.
Abstract
from arXiv · showhide
In spite of the compelling achievements that deep neural networks (DNNs) have made in medical image computing, these deep models often suffer from degraded performance when being applied to new test datasets with domain shift. In this paper, we present a novel unsupervised domain adaptation approach for segmentation tasks by designing semantic-aware generative adversarial networks (GANs). Specifically, we transform the test image into the appearance of source domain, with the semantic structural information being well preserved, which is achieved by imposing a nested adversarial learning in semantic label space. In this way, the segmentation DNN learned from the source domain is able to be directly generalized to the transformed test image, eliminating the need of training a new model for every new target dataset. Our domain adaptation procedure is unsupervised, without using any target domain labels. The adversarial learning of our network is guided by a GAN loss for mapping data distributions, a cycle-consistency loss for retaining pixel-level content, and a semantic-aware loss for enhancing structural information. We validated our method on two different chest X-ray public datasets for left/right lung segmentation. Experimental results show that the segmentation performance of our unsupervised approach is highly competitive with the upper bound of supervised transfer learning.
1 Introduction
Domain shift degrades DNN segmentation performance on new datasets, while supervised transfer learning depends on additional labeled target data. SeUDA addresses this gap through unsupervised, semantic-aware adversarial adaptation.
- Domain shift causes DNN segmentation performance to degrade on datasets acquired with different protocols or clinical centers.
- Supervised transfer learning fine-tunes a source model with labeled target data, but obtaining those annotations is expensive or sometimes infeasible.
- Unsupervised domain adaptation avoids target labels by adapting test data toward the source distribution, including through histogram matching and GAN-based distribution mapping.
- SeUDA transforms target images into source-like images while preserving structural information, allowing the established source segmentation DNN to process them directly.
2 Method
SeUDA separates source-domain segmentation from unsupervised target-to-source image adaptation. Its GAN framework combines distribution matching, cycle consistency, and semantic-aware adversarial learning to preserve clinically relevant structure.
- Image Transformation: SeUDA adapts unlabeled target images toward source image space so the established source segmenter can be applied directly to transformed images.
- Segmentation Network: The segmenter is trained independently from the adaptation GANs, enabling flexible use of a high-performance architecture built with residual, dilated, and multi-scale features.
- Adversarial Learning: A target-to-source generator and source discriminator adversarially match generated images to real source images.
- Cycle-Consistency Learning: Cycle consistency uses inverse generators and an L1 pixel-level penalty to preserve target contents during unpaired image transformation.
- Semantic-Aware Adversarial Learning: The semantic-aware loss adds a discriminator comparing source lung masks with source-segmenter predictions on transformed images, penalizing semantic distortion.
- Training Procedure: The generators and three discriminators are optimized jointly, with α, β, and λ set to 0.5, 10, and 0.5, respectively.
3 Experimental Results
SeUDA was evaluated for lung segmentation across the Montgomery and JSRT chest X-ray datasets under multiple domain-adaptation settings. It improved substantially over direct testing and histogram matching, while preserving image structure and stabilizing transformation.
- Datasets and Metrics: The evaluation used the 138-case Montgomery set and 247-case JSRT set, whose distributions differ in disease type, intensity, and contrast.
- Datasets and Metrics: The study measured Dice, recall, precision, and average surface distance, with lower ASD indicating better boundary performance.
- Comparison of Methods: 95.61% to 79.47%: average Dice dropped after directly applying the source segmenter to target data, while average ASD increased from 2.34 to 11.04 mm.
- Comparison of Methods: 15.04%: SeUDA increased average Dice over T-noDA; versus T-HistM, average Dice increased by 3.97% and average ASD decreased from 5.19 mm to 3.18 mm.
- Qualitative Results: SeUDA produced source-like target appearances while preserving positions, contents, semantic structures, and clinical clues, supporting accurate lung segmentation.
- Semantic-Aware Loss: 0.25% vs. 2.03%: Dice standard deviation was lower for SeUDA than CyUDA across ten initializations, while ASD standard deviation was 0.16 vs. 1.19 mm.
4 Conclusion
SeUDA uses GAN-based image transformation and a three-part objective to adapt medical segmentation across domains. On chest X-ray lung segmentation, it is highly competitive with supervised transfer learning.
- SeUDA transforms target images toward the source domain and applies the established source segmentation DNN directly.
- Its objective combines GAN, cycle-consistency, and semantic-aware losses for distribution mapping, pixel-level content preservation, and structural information enhancement.
- SeUDA is highly competitive with supervised transfer learning for lung segmentation in chest X-rays.