PaperlayerView exact revision on arXiv

Source-linked AI summary

Causality-inspired Single-source Domain Generalization for Medical Image Segmentation

Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, Daniel Rueckert

arXiv:2111.12525v5cs.CV

TL;DR

Medical segmentation models trained on one domain often fail under acquisition-driven shifts, while practical alternatives requiring target or multiple-source data may be unavailable. The paper proposes causality-inspired augmentation that varies appearance and breaks spurious object correlations, yielding consistent gains across three unseen-domain segmentation tasks. It also reports source-domain trade-offs and remaining empirical tuning requirements.

  • Problem

    Single-source domain generalization seeks robust medical image segmentation using only one source domain despite acquisition-driven shifts and unseen target distributions.

  • Method

    The method combines randomly weighted shallow-network appearance transformations with causal intervention that independently resamples appearances of potentially correlated objects.

  • Results

    The approach consistently outperforms peer methods across cross-modality, cross-sequence, and cross-center segmentation scenarios on unseen domains.

  • Takeaways & Limitations

    Causality-inspired augmentation provides a bottom-up strategy for improving robustness when only one source domain is available.

  • Takeaways & Limitations

    GIN layer counts and IPA configurations require empirical choices, while suppressing domain-specific information slightly lowers source-domain performance on abdominal CT and prostate MRI.

Abstract

from arXiv · show

Deep learning models usually suffer from domain shift issues, where models trained on one source domain do not generalize well to other unseen domains. In this work, we investigate the single-source domain generalization problem: training a deep network that is robust to unseen domains, under the condition that training data is only available from one source domain, which is common in medical imaging applications. We tackle this problem in the context of cross-domain medical image segmentation. Under this scenario, domain shifts are mainly caused by different acquisition processes. We propose a simple causality-inspired data augmentation approach to expose a segmentation model to synthesized domain-shifted training examples. Specifically, 1) to make the deep model robust to discrepancies in image intensities and textures, we employ a family of randomly-weighted shallow networks. They augment training images using diverse appearance transformations. 2) Further we show that spurious correlations among objects in an image are detrimental to domain robustness. These correlations might be taken by the network as domain-specific clues for making predictions, and they may break on unseen domains. We remove these spurious correlations via causal intervention. This is achieved by resampling the appearances of potentially correlated objects independently. The proposed approach is validated on three cross-domain segmentation tasks: cross-modality (CT-MRI) abdominal image segmentation, cross-sequence (bSSFP-LGE) cardiac MRI segmentation, and cross-center prostate MRI segmentation. The proposed approach yields consistent performance gains compared with competitive methods when tested on unseen domains.

I. INTRODUCTION

Medical segmentation models degrade under acquisition-driven domain shifts, motivating single-source generalization through causal analysis and simulated interventions. The proposed augmentation varies image appearance and independently resamples potentially correlated objects, with evaluation across three cross-domain settings.

  • Acquisition differences in modalities, scanning protocols, or device manufacturers create domain shifts that reduce performance when testing distributions differ from training.
  • Background objects can spuriously correlate with target objects, causing models to rely on domain-specific clues whose relationships may break in unseen domains.
  • Single-source domain generalization trains with one source domain because UDA and MDG require target-domain or multiple-source data that may be unavailable due to cost or privacy.
  • The approach uses randomly weighted shallow networks for diverse intensity and texture transformations, encouraging reliance on domain-invariant shape information.
  • Causal intervention independently resamples appearances of potentially confounded objects, using spatially variable transformations to mitigate shifted correlations during training.
  • The testing environment covers cross-modality, cross-sequence cardiac MRI, and cross-center prostate MRI segmentation settings with varied anatomical structures.

II. RELATED WORKS

Prior work addresses domain robustness through adaptation, invariant representations, and data augmentation, while causal intervention provides a framework for separating correlations from causal relationships. This paper adapts those ideas to single-source medical image segmentation with task-specific augmentation.

  • UDA uses unlabeled target data, whereas MDG learns domain-invariant features from multiple source domains; both differ from the single-source setting.
  • Single-source methods include removing features associated with large loss gradients, unifying image-feature statistics, and deliberately perturbing training samples.
  • Cutout, Mixup, and RandConv improve robustness through feature occlusion, sample interpolation, or random image alterations that encourage shape-based learning.
  • Medical imaging approaches include photometric and geometric transformations, generative contrast-agnostic MRI synthesis, and adversarial bias-field augmentation for segmentation.
  • Causal intervention fixes a variable of interest while incorporating other variables fairly, distinguishing interventional distributions from observed conditional correlations.
  • The paper builds on causal representation learning and augmentation-based intervention, focusing on an augmentation model tailored to cross-domain medical image segmentation.

III. METHOD

The method simulates acquisition shifts through appearance transformations and intervention on spurious correlations, encouraging segmentation based on invariant anatomical information.

  • Method overview: The approach augments source images with simulated intensity/texture shifts and shifted correlations to improve robustness to acquisition-domain changes.It combines GIN and IPA as additional steps in a standard data-augmentation pipeline.
  • Method overview: GIN transforms training images into diverse appearances using randomly weighted shallow convolutional networks sampled during training.The transformations preserve anatomical shapes while varying intensities and textures.
  • Method overview: IPA independently resamples appearances of potentially confounded background and target objects to remove spurious correlations.It implements this intervention through spatially variable blending of two GIN-augmented versions of the same image.
  • Problem formulation: The causal formulation models image appearance as acquisition-dependent while treating content, representation, and segmentation labels as unaffected by acquisition changes.The method assumes content remains unchanged across source and target domains in the experiments.
  • Training objective: Training applies different photometric transformations to the same image and combines segmentation losses with a distribution-divergence objective.The transformations are sampled from a family at each iteration, and the divergence is measured with Kullback–Leibler divergence.

B. Global intensity non-linear augmentation

GIN uses randomly sampled shallow networks to generate diverse, spatially resolution-preserving intensity and texture transformations while retaining image shape information.

  • GIN design: GIN transforms image appearances with shallow convolutional networks whose weights are randomly sampled at each training iteration.The networks use Leaky ReLU nonlinearities and omit downsampling to maintain spatial resolution.
  • GIN design: GIN applies piece-wise linear functions to images in a spatially invariant manner while preserving anatomical shape information.The transformations operate at the pixel or small local patch level.
  • GIN design: Random convolutional kernels use small receptive fields, while Leaky ReLU layers make the transformations nonlinear.New transformation networks are sampled at each iteration to produce varied appearances.
  • GIN design: The transformed image interpolates the random-network output with the original image using a random coefficient and is renormalized by its Frobenius norm.The interpolation coefficient is sampled from U(0, 1).
  • GIN rationale: Randomly changing intensities and textures discourage stable decision rules based on appearance, encouraging reliance on invariant shape information.The generic transformations also avoid being over-specific to a particular target domain and are computationally efficient as shallow networks.

C. Interventional pseudo-correlation augmentation

Interventional pseudo-correlation augmentation removes acquisition-induced confounding by independently resampling appearances of potentially correlated background objects. IPA implements this intervention approximately by spatially varying blends of randomly transformed images.

  • Motivation: Background objects can provide domain-specific segmentation clues because acquisition processes spuriously correlate them with the object of interest.These correlations may break or shift differently in target domains, reducing out-of-domain robustness.
  • Intervention: Causal intervention do(Xf = xf) removes the confounding path A →Xf and resamples correlated-object appearances independently of Xf.The approach assumes the content factor remains unchanged across domains and focuses on acquisition-factor confounding.
  • IPA implementation: IPA uses a new pseudo-correlation map each iteration to assign different appearance transformations to different pixels or patches.The maps are generated by interpolating along a lattice of randomly valued control points, approximating independent appearance resampling.
  • Practical design: The method addresses unknown correlated objects, multiple entangled background effects, and shortcuts caused by fixing the foreground with ground-truth masks.These practical issues motivate an augmentation-based approximation rather than direct object-level intervention.
  • IPA implementation: The IPA output blends two GIN-transformed versions using pixel-wise coefficients: T1(x; θ1, θ2, b) = gθ1(x) ⊙b + gθ2(x) ⊙(1 −b).Swapping b and 1 −b produces an additional augmented image, while preserving spatial correspondence between pixels and labels.

D. Training objective

Training is end-to-end with a loss combining segmentation terms and a diversity-weighting coefficient. The segmentation output is represented as logits and converted to probabilities with softmax.

  • Training objective: The network is trained end-to-end using Eq. 3, with raw logits converted to probabilities by a softmax function.The segmentation loss combines multi-class cross-entropy and soft Dice loss, and λdiv is set to 10.0.

IV. EXPERIMENTS

The experiments evaluate single-source domain generalization across cross-modality, cross-sequence, and cross-center medical image segmentation settings. Dataset organization and evaluation protocols define source-to-target testing without target-domain training.

  • Scenarios: The study evaluates cross-modality abdominal CT-to-T2-SPIR MRI, cross-sequence cardiac bSSFP-to-LGE MRI, and cross-center prostate MRI segmentation.The prostate setting uses six centers, with one source center and five target centers in each repeated experiment.
  • Datasets: Table I summarizes the datasets and source-target splits used for the three cross-domain segmentation scenarios.The originally 3-D datasets are reformatted to 2-D and resized to 192×192, with modality-specific preprocessing.
  • Results: Table II reports performance when models train on source domains and are evaluated on unseen target domains, with target-domain training and testing provided as upper bounds.The proposed approach consistently outperforms peer methods across the three scenarios.
  • Results: Qualitative results are presented for abdominal CT-MRI, cardiac bSSFP-LGE, and prostate cross-center segmentation, with source-domain images in the rightmost column.The rows correspond to the three cross-domain scenarios.
  • Evaluation protocol: Dice score on a 0-100 scale measures overlap between predictions and ground truth.Source-domain abdominal and prostate data use 70%-10%-20% training, validation, and testing splits, while target images are used for testing.

B. Network architecture and training configurations

The method uses a U-Net-based segmentation network and an end-to-end augmentation training procedure combining GIN and IPA. Comparisons include ERM and recent single-source domain generalization methods.

  • Network architecture and training configurations: The segmentation network is a U-Net with an EfficientNet-b2 backbone, trained with Adam from an initial learning rate of 3 × 10^-4.Learning-rate decay is applied, and evaluation occurs at the 2k-th epoch when the learning rate reaches zero.
  • Baselines: The method is compared with ERM, Cutout, RSC, MixStyle, AdvBias, and the closely related RandConv approach.These baselines target robustness through patch removal, feature removal, feature-statistic mixing, adversarial perturbations, or randomized convolutions.
  • End-to-end training: Each training iteration samples a source image-label pair, samples GIN intensity/texture transformations, computes a pseudo-correlation map, and generates augmented images with IPA.The training loss is then computed using Eq. 3 and network parameters are updated.
  • Results: The proposed approach consistently outperforms peer methods across three cross-domain scenarios.Its gains over RandConv suggest more effective simulation of domain shifts and stronger robustness on unseen domains.
  • Feature-space analysis: Target-domain t-SNE features from the proposed method remain close to source features for the same class and separated across different classes.The visualization uses features collected from the last hidden layer of abdominal segmentation networks.

1) Configurations of GIN:

The study evaluates how GIN performance changes with convolutional-layer depth and hidden-layer channel count, using abdominal CT-MRI segmentation with IPA disabled. The authors motivate moderate depth as a balance between insufficient nonlinearity and overly aggressive augmentation.

  • Configurations of GIN:: Figure 6 visualizes source-versus-target feature spaces for networks trained with different domain-generalization techniques in abdominal CT-MR segmentation.Suffixes S and T identify source CT and target MRI features, respectively.
  • Configurations of GIN:: Too many convolutional layers may produce unrealistically aggressive augmentations that deviate from reality.
  • Configurations of GIN:: GIN configurations vary the number of convolutional layers and hidden-layer channels to assess their effects on abdominal CT-MRI segmentation.The default setting is 4 layers and 2 hidden-layer channels; IPA is disabled for this analysis.
  • Configurations of GIN:: The layer-count experiments support the authors’ intuition about how convolutional depth affects GIN, while channel effects are harder to predict because GIN is nonlinear.

2) Configurations of IPA:

The IPA ablation examines whether independently resampling appearances of potentially correlated objects improves domain generalization. Results support IPA’s benefit, particularly in cardiac and prostate settings, while the broader method retains practical limitations and possible extensions.

  • Configurations of IPA:: Removing IPA in ablation studies validates the benefit of mitigating shifted-correlation effects, especially for cardiac and prostate segmentation.The reported Wilcoxon signed-rank tests give pvalues < 0.001 for these settings.
  • Configurations of IPA:: Randomly sampled superpixels provide an alternative pseudo-correlation-map design evaluated in Fig. 7 and the last row of Table IV.
  • Configurations of IPA:: The superpixel-based design may be sub-optimal because superpixels often coincide with ground-truth masks, making foreground features unnatural shortcuts.
  • Discussion and Conclusion: The approach consistently improves performance across three scenarios but slightly lowers source-domain testing performance on abdominal CT and prostate MRI versus ERM.The reported downgrades are approximately -2.0/100 for abdominal CT and -1.0/100 for prostate MRI.
  • Discussion and Conclusion: GIN and IPA still require empirical hyper-parameter choices, and future work could combine the method with multi-source generalization or target anatomical-shape shifts.
Loading 2111.12525v5…