PaperlayerView exact revision on arXiv

Source-linked AI summary

Realistic Adversarial Data Augmentation for MR Image Segmentation

Chen Chen, Chen Qin, Huaqi Qiu, Cheng Ouyang, Shuo Wang, Liang Chen, Giacomo Tarroni, Wenjia Bai, Daniel Rueckert

arXiv:2006.13322v1eess.IVcs.CVcs.LG

TL;DR

Deep-learning medical image segmentation generally requires large labelled datasets, but collecting and manually labelling them is costly and can be impractical. The method adversarially generates realistic multiplicative MR bias-field transformations and uses prediction consistency regularization without generative networks. In ACDC low-shot segmentation, the method outperformed all supervised baselines and beat VAT semi-supervised with one labelled subject (average Dice 0.686 vs 0.625).

  • Problem

    Deep-learning medical image segmentation generally requires large labelled datasets, but collecting and manually labelling them is costly and can be impractical.

    Source: §1 · p. 1

  • Method

    The method adversarially generates realistic multiplicative MR bias-field transformations and uses prediction consistency regularization without generative networks.

    Source: §1 · p. 2, §2.2 · pp. 4–5

  • Results

    In ACDC low-shot segmentation, the method outperformed all supervised baselines and beat VAT semi-supervised with one labelled subject (average Dice 0.686 vs 0.625).

    Source: §3.2 · pp. 6–7

  • Takeaways & Limitations

    The approach uses explainable, controllable parameters and can be extended to multi-class segmentation or used as a plug-in in general segmentation networks.

    Source: §4 · p. 9

Abstract

from arXiv · show

Neural network-based approaches can achieve high accuracy in various medical image segmentation tasks. However, they generally require large labelled datasets for supervised learning. Acquiring and manually labelling a large medical dataset is expensive and sometimes impractical due to data sharing and privacy issues. In this work, we propose an adversarial data augmentation method for training neural networks for medical image segmentation. Instead of generating pixel-wise adversarial attacks, our model generates plausible and realistic signal corruptions, which models the intensity inhomogeneities caused by a common type of artefacts in MR imaging: bias field. The proposed method does not rely on generative networks, and can be used as a plug-in module for general segmentation networks in both supervised and semi-supervised learning. Using cardiac MR imaging we show that such an approach can improve the generalization ability and robustness of models as well as provide significant improvements in low-data scenarios.

1 Introduction

Medical image segmentation depends on large labelled datasets, which are difficult and costly to collect. The paper proposes adversarial augmentation using realistic MR bias-field transformations and evaluates it in low-data settings.

1 Introduction

  • Deep-learning segmentation needs large labelled datasets, but collecting, sharing, and manually labelling them is challenging and costly.The introduction frames the need for augmentation around these data-acquisition constraints.Source: §1 · p. 1
  • Conventional augmentation uses random intensity or geometric transformations, while generative methods can be difficult to train and may suffer mode collapse.Generative methods learn transformations from real-world data distributions, but are sensitive to hyperparameter tuning.Source: §1 · p. 2
  • The proposed method adversarially generates realistic MR bias-field transformations to improve generalization and robustness without generative networks.The authors describe the generated examples as realistic, challenging perturbations intended to encourage robust semantic features.Source: §1 · p. 2
  • The framework is a plug-in for general segmentation networks, supports supervised and semi-supervised learning, and outperforms competitive baselines in challenging low-data cardiac MR settings.The contribution passage reports evaluation on a public cardiac MR segmentation dataset.Source: §1 · p. 2

2 Adversarial Data Augmentation with Robust Optimization

The method extends virtual adversarial training by perturbing MR images with smooth, multiplicative bias fields rather than pixel-wise noise. It searches for fields that challenge the segmentation network, then regularizes the network to make predictions consistent under those perturbations.

2.1 Virtual Adversarial Training

  • VAT constructs adversarial examples by maximizing prediction divergence under a small input-noise perturbation, then uses them to regularize the network.The perturbation is generated from the gradient of KL divergence with respect to a random noise vector.Source: §2.1 · p. 3

2.2 Adversarial Training by Modelling Intensity Inhomogeneities

  • A bias field is a smooth, low-frequency multiplicative intensity map modeled from grid control points and interpolated with a third-order B-spline.The transformation multiplies image intensity by the bias field.Source: §2.2 · p. 4
  • Rather than sampling random fields, the method searches for control-point values that maximize the prediction distance, then updates network parameters to minimize that distance.Control points are updated by projected gradient ascent in log space to produce positive bias fields.Source: §2.2 · pp. 4–5
  • The composite distance combines KL divergence with a contour-based loss that captures boundary mismatch.The contour term compares Sobel-derived edges of foreground predictions.Source: §2.2 · p. 5
  • Because adversary construction uses network predictions rather than manual labels, the consistency regularizer can use labelled data in supervised learning and labelled plus unlabelled data in semi-supervised learning.The semi-supervised objective adds a consistency term for unlabelled examples to the supervised objective.Source: §2.2 · p. 5

3 Experiments

The experiments evaluate adversarial bias-field augmentation for cardiac myocardium segmentation in low-data and cross-population settings. The method improves results over augmentation baselines in both settings, including when training uses only normal cases.

3.1 Dataset and Experiment Settings

  • The evaluation used ACDC cardiac MR data and tested low-shot learning with one or three labelled subjects in supervised and semi-supervised settings.The low-shot experiments used five different labelled sets for each condition.Source: §3.1 · p. 5, §3 · p. 5, §3.2 · p. 6

3.2 Experiments and Results

  • With only one or three labelled subjects, Adv Bias outperformed all baselines in supervised learning and outperformed VAT in semi-supervised learning.Its semi-supervised result was competitive with cGANs; cGANs was evaluated only in the semi-supervised setting.Source: Table 1 · p. 7, §3.2 · p. 7
  • Training on normal cases, Adv Bias outperformed Mixup and VAT with consistent improvements across five test populations.The results also report improvement on the challenging MINF population.Source: Table 2 · p. 7, §3.2 · p. 8
  • Adversarial bias fields increased Dice from 0.852 to 0.868, while Dcomp regularization increased average Dice from 0.859 to 0.868 versus DKL alone.The authors report that adversarial attacks target challenging, uncertain images, unlike random bias-field augmentation.Source: §3.2 · p. 8

4 Discussion and Conclusion

The paper presents realistic adversarial data augmentation for medical image segmentation and reports improved generalization and robustness in cardiac segmentation. It uses controllable parameters to model MR appearance variation without generative networks.

4 Discussion and Conclusion

  • The authors report that modeling bias fields and adversarial learning promotes robust semantic features for cardiac image segmentation and alleviates data scarcity in low-data and cross-population experiments.The abstracted conclusion links the demonstrated results to improved generalization and robustness.Source: §4 · pp. 8–9
  • The method uses a small set of explainable, controllable parameters to add MR-realistic image appearance variations without generative networks.The paper describes the augmentation as complementary to conventional methods.Source: §4 · p. 9
  • The approach is described as extendable to multi-class segmentation and usable in general segmentation networks.The authors present it as a way to improve model generalization and robustness.Source: §4 · p. 9

Supplementary material: Realistic Adversarial Data Augmentation for MR Image Segmentation

The supplementary material compares adversarial and random bias-field attacks and reports segmentation results across populations. The population comparison highlights performance on unseen pathological cases after training on normal cases only.

Supplementary material: Realistic Adversarial Data Augmentation for MR Image Segmentation

  • The supplementary figure compares performance for adversarial bias-field attack and random bias-field attack.The passage identifies the two approaches but provides no numerical outcome.Source: Supplementary material: Realistic… · p. 12
  • On unseen pathological cases, the method produced more accurate myocardium segmentation despite training only on normal cases.The figure shows subject-level Dice scores, with color indicating each subject's group.Source: Supplementary material: Realistic… · p. 12
  • The supplementary tables compare Dcomp with DKL and random with adversarial bias fields, reporting average scores across 84 subjects from five populations.The listed metrics are Dice, Hausdorff distance, and VolumeSim.Source: Supplementary material: Realistic… · p. 13

Related papers

Loading 2006.13322v1…