Source-linked AI summary
Enhancing Pseudo Label Quality for Semi-Supervised Domain-Generalized Medical Image Segmentation
Huifeng Yao, Xiaowei Hu, Xiaomeng Li
TL;DR
Medical image segmentation must generalize across unseen domains despite expensive annotation and limited applicability of methods requiring fully labeled data or domain labels. The paper introduces Fourier-based augmentation and confidence-aware cross pseudo supervision to improve pseudo labels for unlabeled images from unknown domains. On M&Ms, it reports an 11.67% Dice advantage over prior art using domain labels with 2% labeled data.
Problem
Medical image segmentation faces domain shifts, while existing approaches may require fully labeled source domains or source-domain labels.
Method
The method combines Fourier-based cross-domain augmentation with confidence-aware cross pseudo supervision for unlabeled images from unknown domains.
Results
11.67% Dice improvement over prior art using domain labels was achieved on M&Ms with 2% labeled data.
Takeaways & Limitations
The method achieves state-of-the-art performance on the M&Ms and SCGM public datasets.
Abstract
from arXiv · showhide
Generalizing the medical image segmentation algorithms to unseen domains is an important research topic for computer-aided diagnosis and surgery. Most existing methods require a fully labeled dataset in each source domain. Although some researchers developed a semi-supervised domain generalized method, it still requires the domain labels. This paper presents a novel confidence-aware cross pseudo supervision algorithm for semi-supervised domain generalized medical image segmentation. The main goal is to enhance the pseudo label quality for unlabeled images from unknown distributions. To achieve it, we perform the Fourier transformation to learn low-level statistic information across domains and augment the images to incorporate cross-domain information. With these augmentations as perturbations, we feed the input to a confidence-aware cross pseudo supervision network to measure the variance of pseudo labels and regularize the network to learn with more confident pseudo labels. Our method sets new records on public datasets, i.e., M&Ms and SCGM. Notably, without using domain labels, our method surpasses the prior art that even uses domain labels by 11.67% on Dice on M&Ms dataset with 2% labeled data. Code is available at https://github.com/XMed-Lab/EPL_SemiDG.
Introduction
Medical image segmentation must handle domain shifts while reducing costly annotation requirements. This paper proposes confidence-aware cross pseudo supervision with Fourier-based augmentation for semi-supervised domain generalization without domain labels.
- Medical images exhibit domain shifts from patient populations, scanners, and acquisition settings, degrading performance on unseen domains.
- Domain generalization trains on source domains and tests on an unseen target domain without using target-domain information.
- Existing semi-supervised domain-generalization methods may require source-domain labels, which are difficult to obtain in clinical practice.
- The proposed method uses Fourier transformation to capture cross-domain low-level statistics and augment unlabeled images with cross-domain information.
- Confidence-aware cross pseudo supervision measures pseudo-label variance across original and Fourier-augmented images to improve pseudo-label quality.
- 11.67% Dice improvement over prior art using domain labels was achieved on M&Ms with 2% labeled data.
Related Work
Related work spans representation learning, learning-strategy design, data manipulation, and semi-supervised segmentation. The paper adapts Fourier augmentation and cross pseudo supervision to unlabeled images from out-of-domain distributions.
- Domain-generalization methods include representation learning, learning-strategy design, and data manipulation.
- Fourier-based methods have been used for domain generalization by exchanging amplitude spectra across source-domain images.
- This paper uses Fourier transformation to obtain cross-domain low-level statistics and augment unlabeled images, then measures pseudo-label confidence.
- Semi-supervised segmentation commonly uses self-training, consistency learning, or combinations of both.
- Unlike prior cross pseudo supervision with same-domain unlabeled images, this method addresses unlabeled images from out-of-domain distributions.
Methodology
The method augments unlabeled images with cross-domain low-level information, then uses two differently initialized segmentation networks and confidence-aware cross pseudo supervision to weight pseudo-label training by prediction reliability.
- Data Augmentation by Fourier Transformation: Fourier transformation augments source images by exchanging central low-frequency amplitude information from another sample, producing an image with cross-domain low-level statistics.The merged frequency-domain sample is transformed back to the image domain before segmentation.
- Confidence-Aware Cross Pseudo Supervision: Two same-structure segmentation networks with different initializations process both original and transformed images to generate complementary predictions.The predictions are obtained separately for each network and image version.
- Confidence-Aware Cross Pseudo Supervision: Cross pseudo supervision uses predictions from one network to supervise the other on unlabeled data, addressing the lack of ground-truth supervision across multiple domains.The approach differs from same-domain semi-supervised segmentation because unlabeled data here may exhibit larger variation.
- Confidence-Aware Cross Pseudo Supervision: Prediction variance between original and transformed images is measured to estimate pseudo-label quality, with larger differences indicating lower confidence.The variance is computed from the discrepancy between the two predictions, described as a KL-divergence.
- Confidence-Aware Cross Pseudo Supervision: The confidence-aware loss combines cross pseudo-supervision losses with supervised Dice loss, while β balances the two components.At test time, the ensemble of the two models’ predictions is used as the final result.
Experimental Results
The method is evaluated on M&Ms and SCGM against established segmentation and domain-generalization baselines, with results reported using Dice and visual comparisons. It achieves the best reported M&Ms performance across settings, improves SCGM Dice by approximately 7% over the previous best method, and ablations support both proposed components.
- Datasets and evaluation: The evaluation uses M&Ms and SCGM, which represent multi-domain medical imaging collected across scanners, clinical centres, or acquisition systems.M&Ms includes four vendor-defined domains, while SCGM includes four centre-defined domains.
- Datasets and evaluation: Dice (%) is the evaluation metric, comparing predicted segmentation P with ground-truth image G through their overlap.A larger Dice score indicates a better segmentation result.
- Comparison methods: The comparisons include nnUNet, SDNet+Aug, LDDG, SAML, and Meta, spanning standard segmentation, augmentation, feature-learning, and meta-learning baselines.LDDG is described as fully supervised, whereas Meta is described as semi-supervised domain-generalized medical image analysis.
- M&Ms results: ≈6% and ≈12% Dice improvements are reported on M&Ms with 5% and 2% labeled data, respectively, over the previous best method.The method achieves the best performance across the different M&Ms settings, including evaluation on domain A after training on domains B, C, and D.
- SCGM results: ≈7% Dice improvement is reported on SCGM with 20% labeled data compared with the previous best method.Visual results are reported as consistent with ground truth across all four SCGM domains.
- Ablation study: Each proposed component gives a clear improvement in the M&Ms ablation, while the combined method produces pseudo labels more consistent with ground truth than CPS.The ablation compares CPS, Fourier-transform-based augmentation with CPS, and confidence-aware cross pseudo supervision.
Conclusion
The paper presents Fourier-transform-based data augmentation and confidence-aware cross pseudo supervision for semi-supervised domain generalization in medical image segmentation. These techniques incorporate cross-domain information, improve pseudo-label quality, and outperform other state-of-the-art methods on two public benchmarks.
- Fourier-transform-based data augmentation obtains low-level statistics from different domains and incorporates cross-domain information into image samples.
- Confidence-aware cross pseudo supervision measures variance between original and Fourier-augmented samples to improve pseudo-label quality and facilitate framework learning.
- The method was evaluated against various methods on two public benchmark datasets and showed superiority over other state-of-the-art methods.