Source-linked AI summary
Rethinking Data Augmentation for Single-source Domain Generalization in Medical Image Segmentation
Zixian Su, Kai Yao, Xi Yang, Qiufeng Wang, Jie Sun, Kaizhu Huang
TL;DR
Single-source domain generalization in medical image segmentation must address unseen-domain shifts using only source data, while prior global-only or random augmentation may provide insufficient diversity and informativeness. The paper proposes class-level location-scale augmentation with saliency-balancing fusion, and reports superior performance in two segmentation tasks. Its formulation rests on representing unseen target data through class-level random variables.
Problem
Single-source domain generalization seeks robust prediction on unseen domains from one source, while existing augmentation may not provide sufficient diversity and informativeness.
Source: Introduction · p. 1Method
The method combines global and class-level location-scale augmentation with saliency-balancing fusion that uses gradient information to guide augmentation.
Source: Introduction · p. 2, Main Methodology · p. 3Results
SLAug achieves superior performance in two segmentation tasks, and saliency-balancing fusion achieves lower source risk and a tighter generalization bound in the reported setting.
Source: Introduction · p. 2Takeaways & Limitations
The paper concludes that global and local augmentation combined with saliency-balancing fusion provides a strategy for single-source medical-image segmentation generalization.
Source: Introduction · p. 2The approach assumes unseen target data can be represented as a linear combination of C class-level random variables following location-scale distributions.
Source: Introduction · p. 2
Abstract
from arXiv · showhide
Single-source domain generalization (SDG) in medical image segmentation is a challenging yet essential task as domain shifts are quite common among clinical image datasets. Previous attempts most conduct global-only/random augmentation. Their augmented samples are usually insufficient in diversity and informativeness, thus failing to cover the possible target domain distribution. In this paper, we rethink the data augmentation strategy for SDG in medical image segmentation. Motivated by the class-level representation invariance and style mutability of medical images, we hypothesize that unseen target data can be sampled from a linear combination of $C$ (the class number) random variables, where each variable follows a location-scale distribution at the class level. Accordingly, data augmented can be readily made by sampling the random variables through a general form. On the empirical front, we implement such strategy with constrained B$\acute{\rm e}$zier transformation on both global and local (i.e. class-level) regions, which can largely increase the augmentation diversity. A Saliency-balancing Fusion mechanism is further proposed to enrich the informativeness by engaging the gradient information, guiding augmentation with proper orientation and magnitude. As an important contribution, we prove theoretically that our proposed augmentation can lead to an upper bound of the generalization risk on the unseen target domain, thus confirming our hypothesis. Combining the two strategies, our Saliency-balancing Location-scale Augmentation (SLAug) exceeds the state-of-the-art works by a large margin in two challenging SDG tasks. Code is available at https://github.com/Kaiseem/SLAug .
Introduction
The paper targets single-source domain generalization (SDG), where medical-image domain shifts make training-data diversity and informativeness important. It proposes class-aware location-scale augmentation with saliency-balancing fusion and reports superior performance in two segmentation tasks.
Introduction
- Single-source domain generalization seeks robust prediction on unseen domains using only source-domain data, making augmentation a key way to increase training diversity and informativeness.Prior approaches may use extra data or require diverse training data; global-only and unconstrained random augmentation have limitations for medical-image segmentation.Source: Introduction · p. 1
- Class-level representation invariance, mutable style, and pixel-level labels motivate transforming class regions separately and combining them into diverse samples.The examples contrast similar kidney contrast tendencies across CT and MRI with differing liver appearance; masks can crop class-level regions for augmentation.Source: Figure 1 · p. 1, Introduction · p. 2
- The method hypothesizes that unseen target data can be represented as a linear combination of C class-level random variables following location-scale distributions.Location-scale augmentation samples augmented data from this general form.Source: Introduction · p. 2
- Global and local location-scale augmentation use constrained B´ezier transformations, while saliency-balancing fusion uses gradient information to guide augmentation orientation and magnitude.The fusion strategy preserves large-gradient regions from global samples and combines them with local samples to avoid samples too close to the source or too far out-of-distribution.Source: Introduction · p. 2
- SLAug achieves superior performance in two segmentation tasks, demonstrating generalization capability.The paper also reports that saliency-balancing fusion lowers source risk and tightens the generalization bound.Source: Introduction · p. 2
Related Work
Prior work addresses domain generalization through varied augmentation, mask-based sample construction, and saliency analysis. The paper frames its approach as adding class-level information to augmentation for medical-image SDG.
Related Work
- Existing SDG methods use different generation strategies, while the paper introduces additive class-level information into augmentation for medical-image generalization.The related-work discussion also cites medical-image methods using augmentation to remove spurious correlations and dual normalization for cross-modality generalization.Source: Related Work · p. 2
- Mask-based methods cut and paste predicted classes or replace masked objects, but the cited approaches are not designed for medical SDG.The paper distinguishes these approaches from augmentation tailored to medical segmentation under domain generalization.Source: Related Work · p. 2
- The method detects saliency using gradients from a pretrained model, without modifying the network or requiring additional annotation.The discussion describes gradient-based saliency as measuring each pixel's impact on the classification result.Source: Related Work · p. 3
Main Methodology
SLAug combines location-scale augmentation with saliency-balancing fusion. Its augmentation operates globally and on class-specific image regions, using segmentation masks to decouple those regions.
Main Methodology
- SLAug consists of Location-scale Augmentation and Saliency-balancing Fusion.Figure 2 presents the framework workflow.Source: Main Methodology · p. 3
Preliminary
- The preliminary formulation decomposes an image into class-specific regions using binary masks, with the regions summing to the original image.Each pixel belongs to one class, so the masks collectively cover the image.Source: Preliminary · p. 3
Location-scale Augmentation
- Global Location-scale Augmentation shifts the image distribution, while Local Location-scale Augmentation transforms class-level regions to explore diverse or extreme unseen-domain appearances.LLA combines its augmented class regions linearly.Source: Location-scale Augmentation · p. 3
- The method uses a constrained, smooth monotonic B´ezier mapping for medical images because absolute intensity values convey information about underlying substructures.Its mapping is bounded by the lowest and highest values in the non-blank image region; GLA and LLA use location-scale factors, and LLA applies transformations to foreground regions.Source: Location-scale Augmentation · p. 3
A. Global LocaƟon-scale AugmentaƟon (GLA)
The section combines global and class-level location-scale augmentation with saliency-guided fusion. It also gives a bound on unseen-domain risk under stated assumptions.
A. Global LocaƟon-scale AugmentaƟon (GLA)
- GLA and LLA use location-scale factors for global and class-specific transformations, with foreground-only nonlinear operations and constrained factor variation.The described settings use a small scale factor to keep samples near the source and constrain the shift to neither be too small nor too large.Source: A. Global LocaƟon-scale AugmentaƟon… · p. 4, Figure 3 · p. 4
Theoretical Evidence
- The theorem bounds single-source domain-generalization risk by σ + δ, where σ bounds empirical source risk and δ bounds source-distribution divergence.The proof applies when the unseen target distribution is represented within the source-mixture convex hull; the bound depends only on σ and δ.Source: Theoretical Evidence · p. 4
Saliency-balancing Fusion
- Saliency-balancing Fusion uses gradients from GLA-augmented images to retain large-gradient regions while replacing other regions with LLA-augmented content.The fused examples combine both augmentations, and the training pipeline uses GLA-augmented and fused images.Source: Saliency-balancing Fusion · pp. 4–5
Experiments and Results
SLAug is evaluated on abdominal cross-modality and cardiac cross-sequence segmentation, with shared preprocessing and training settings. It outperforms prior methods and ablations indicate that its augmentation components work best together.
Datasets and Preprocessing
- The experiments use abdominal cross-modality and cardiac cross-sequence datasets, with common augmentations applied to all compared methods.The compared methods, including ERM and Supervised, share the same common augmentations for fair comparison.Source: Datasets and Preprocessing · p. 5
Network Architecture and Training Configurations
- The segmentation network is U-Net with an EfficientNet-b2 backbone, trained from scratch and evaluated at epoch 2,000.The grid size is 3 for abdominal and 18 for cardiac data; training uses Adam with batch size 32.Source: Network Architecture and Training… · p. 5
Results and Comparative Analysis
- In cardiac cross-sequence experiments, SLAug improves over prior best performance by 1.68 and 0.43 across the two directions.The paper reports steady performance across the smaller and larger domain-gap situations.Source: Results and Comparative Analysis · p. 5
- 47.77% average improvement over CSDG across four experiments; SLAug achieves average abdominal Dice scores of 88.63 and 83.05 in the two directions.The comparison covers abdominal cross-modality and cardiac cross-sequence tasks.Source: Results and Comparative Analysis · p. 5, Table 1 · p. 6
- Figure 4 compares abdominal CT–MRI and cardiac bSSFP–LGE cases; SLAug generates few misclassified predictions in unseen target domains.The first two columns show source and target images, illustrating appearance shifts.Source: Results and Comparative Analysis · p. 6, Figure 4 · p. 6
Analytical Experiments
- GLA or LLA alone improves performance, but each boosts only one of the two tasks; combining them can outperform most state-of-the-art methods.Adding SBF to the variants yields the best performance, with the three components reported as mutually promoting and indispensable.Source: Analytical Experiments · pp. 6–7, Table 2 · p. 7
- Figure 5 visualizations indicate that SBF combines sensitive GLA regions with larger augmentation magnitude elsewhere.The reported saliency maps show larger augmentation intensity for LLA and identify large-gradient areas that are misclassified in GLA predictions.Source: Figure 5 · p. 7, Analytical Experiments · p. 7
- On the abdominal task, random fusion improves Dice by 0.2 over no fusion, while SBF exceeds random fusion by 2.9 Dice.SBF also improves source performance by 0.8 versus no fusion and 4.0 versus ERM.Source: Figure 6 · p. 7, Analytical Experiments · p. 7
Conclusion
The paper proposes a plug-and-play augmentation strategy for single-source domain generalization in medical image segmentation. It reports superior performance on two challenging tasks and provides theoretical evidence and visualizations.
Conclusion
- The proposed augmentation strategy targets single-source domain generalization in medical image segmentation.The conclusion presents it as a novel strategy for this task.Source: Conclusion · p. 7
- Jointly integrating location-scale augmentation and saliency-balancing fusion yields a plug-and-play module with superior performance on two challenging tasks.The paper also provides theoretical evidence and visualization examples.Source: Conclusion · p. 7
- The authors offer theoretical evidence and visualization examples toward more principled robust-learning techniques in medical fields.Source: Conclusion · p. 7
Additive Experimental Details
Experiments evaluate cross-modality abdominal and cross-sequence cardiac segmentation, with shared preprocessing and common augmentations across comparison methods. Training synchronizes augmentation states for the global and local images so they can be fused while retaining their appearance differences.
Additive Experimental Details
- The evaluation covers abdominal CT and T2-SPIR MRI, plus cardiac bSSFP and LGE MRI, with source-domain splits and target-domain testing described for each dataset.The abdominal dataset has 30 CT and 20 MRI volumes; the cardiac dataset has 45 volumes per sequence.Source: Additive Experimental Details · p. 9
- All 3D volumes are spatially normalized, reformatted into 2D slices, and cropped/resized to 192 × 192 on the XY plane; the proposed approach is an additional stage followed by common augmentations applied to all methods.Source: Additive Experimental Details · p. 9
- GLA and LLA images for a slice use the same random states, aligning geometric transformations while preserving location-scale appearance differences.Shared affine and elastic transformations prevent spatial mismatch during fusion; shared intensity augmentations maintain the appearance difference.Source: Additive Experimental Details · p. 9
Hyper-parameter Analysis
Performance remains steady as grid size changes in the two experiments. Grid size is also discussed in relation to class spacing and fusion-map resolution.
Hyper-parameter Analysis
- Performance remains steady across changes in grid size g in both experiments.At grid size 1, the saliency map is uniformly 1, so only GLA images contribute to the fused image.Source: Hyper-parameter Analysis · p. 10
- For closely spaced classes, a larger grid size is suggested to provide a higher-resolution fusion map.The discussion contrasts cardiac substructures, which are closer together, with separate abdominal organs.Source: Hyper-parameter Analysis · p. 10
More Visualization
The supplementary figures illustrate Bézier transformations and provide additional comparative segmentation visualizations for abdominal and cardiac tasks.
More Visualization
- Figure S8 demonstrates Bézier transformations on normalized images, including an intensity-inversion example.The endpoints are set to (0, 0) and (1, 1), except for the inversion example, which uses (0, 1) and (1, 0).Source: More Visualization · p. 10, Figure S8 · p. 11
- Figure S9 compares source and target images, comparative methods, SLAug, and ground truth across abdominal MRI-CT and cardiac LGE-bSSFP segmentation.Its columns show source images first, target images second, other methods next, SLAug ninth, and ground truth last.Source: More Visualization · p. 10, Figure S9 · p. 11