PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Semi-Supervised Learning for COVID Region Segmentation in Chest CT using Multi-National Data from China, Italy, Japan

Dong Yang, Ziyue Xu, Wenqi Li, Andriy Myronenko, Holger R. Roth, Stephanie Harmon, Sheng Xu, Baris Turkbey, Evrim Turkbey, Xiaosong Wang, Wentao Zhu, Gianpaolo Carrafiello, Francesca Patella, Maurizio Cariati, Hirofumi Obinata, Hitoshi Mori, Kaku Tamura, Peng An, Bradford J. Wood, Daguang Xu

arXiv:2011.11750v1eess.IVcs.CV

TL;DR

The paper addresses cross-site domain shift and constraints on sharing patient data and obtaining annotations for COVID-19 CT analysis. It proposes federated semi-supervised learning for COVID-19 region segmentation, reporting improved generalizability across client databases and an unseen domain after joint training with supervised and unlabeled clients.

  • Problem

    Cross-site data-distribution differences, restricted sharing of sensitive data, and uneven annotation availability challenge robust learning-based COVID-19 image analysis.

    Source: §1 · p. 3

  • Method

    The framework combines federated and semi-supervised learning to train COVID-19 region segmentation models using labeled and unlabeled data across institutions.

    Source: §1 · p. 3

  • Results

    Joint training with supervised and unlabeled clients improved generalizability on each client’s database and an unseen data domain.

    Source: §6 · p. 16

  • Takeaways & Limitations

    The framework preserves patient privacy without requiring database sharing, and semi-supervision potentially reduces annotation burden in distributed settings.

    Source: §1 · p. 3

    The model is trained on COVID-19 cases to segment affected regions, not to distinguish them from other abnormalities; that distinction requires additional data and annotations.

    Source: §6 · p. 16

Abstract

from arXiv · show

The recent outbreak of COVID-19 has led to urgent needs for reliable diagnosis and management of SARS-CoV-2 infection. As a complimentary tool, chest CT has been shown to be able to reveal visual patterns characteristic for COVID-19, which has definite value at several stages during the disease course. To facilitate CT analysis, recent efforts have focused on computer-aided characterization and diagnosis, which has shown promising results. However, domain shift of data across clinical data centers poses a serious challenge when deploying learning-based models. In this work, we attempt to find a solution for this challenge via federated and semi-supervised learning. A multi-national database consisting of 1704 scans from three countries is adopted to study the performance gap, when training a model with one dataset and applying it to another. Expert radiologists manually delineated 945 scans for COVID-19 findings. In handling the variability in both the data and annotations, a novel federated semi-supervised learning technique is proposed to fully utilize all available data (with or without annotations). Federated learning avoids the need for sensitive data-sharing, which makes it favorable for institutions and nations with strict regulatory policy on data privacy. Moreover, semi-supervision potentially reduces the annotation burden under a distributed setting. The proposed framework is shown to be effective compared to fully supervised scenarios with conventional data sharing instead of model weight sharing.

1. Introduction

Chest CT can reveal COVID-19 patterns, and delineating affected regions may help track disease progression, but cross-site data and annotation constraints complicate robust model development. The paper proposes federated semi-supervised learning for COVID-19 region segmentation across institutions.

1. Introduction

  • Chest CT is being investigated for visual patterns of COVID-19, including infiltrates that are usually bilateral, multilobar, and peripheral.The authors note that CT specificity for diagnosis lacks consensus and that ACR guidelines do not recommend it for that purpose.Source: §1 · p. 2
  • Quantifying and characterizing infiltrates may help understand and track disease progression, while manual detection and delineation can be tedious and time-consuming.Deep-learning methods are being developed to facilitate analysis of medical-image findings.Source: §1 · pp. 2–3
  • The proposed federated semi-supervised framework uses diverse data sources to address cross-site domain shift and annotation availability while avoiding sensitive-data sharing.Federated learning shares model weights and gradients rather than patient data, while semi-supervision can use unlabeled data at sites with limited annotations.Source: Figure 1 · p. 2, §1 · p. 3
  • The study evaluates segmentation of COVID-19-related abnormal regions, a time-consuming task that often requires slice-by-slice delineation.The authors describe the framework as benefiting from multiple heterogeneous sources and as adaptable to other deep-learning tasks.Source: §1 · p. 3

2. Related Work

Prior work spans federated learning for privacy-preserving multi-institutional modeling and semi-supervised methods that use unlabeled data. This paper focuses on global semi-supervision: clients with different annotation availability jointly contribute to federated training for COVID-19 region segmentation.

2. Related Work

  • Federated learning enables institutions to use datasets across sites without centralizing or sharing training data.Prior work applied it to EHR, wearable-health, fMRI, and brain-tumor imaging tasks.Source: §2 · p. 3
  • Semi-supervised learning uses labeled and unlabeled data; prior approaches include consistency regularization, teacher-student models, auxiliary tasks, and co-training.Medical-imaging examples include cardiac and skin-lesion segmentation, brain-lesion and left-atrium segmentation, and multi-view volumetric segmentation.Source: §2 · pp. 3–4
  • Federated semi-supervision adds a multi-client dimension: annotation availability can differ across clients, which the authors call “global semi-supervision.”This differs from “local semi-supervision,” which may be used within an individual client; client-level networks can also use other semi-supervised techniques.Source: §2 · p. 4
  • The paper proposes federated semi-supervised training that uses unlabeled data for COVID-19 region segmentation across heterogeneous institutions without sharing sensitive information.The authors describe the existing literature on federated semi-supervised or self-learning for unlabeled medical imaging as limited.Source: §2 · pp. 4–5

3. Coronavirus Affected Region Segmentation

The section frames COVID-19 region segmentation as delineating varied infiltrates in chest CT and uses a 3D U-shape FCN to predict affected regions. It describes the model’s training and inference setup.

3. Coronavirus Affected Region Segmentation

  • COVID-19 infiltrates range from diffuse ground-glass opacity to focal nodules across the illustrated CT scans.Green contours mark the regions annotated by radiologists.Source: Figure 2 · p. 5
  • The FCN uses skip connections to pass earlier feature maps to later layers, and the framework is independent of the exact network architecture.The 3D encoder can be initialized from a pretrained 2D ResNet by converting 2D 3×3 convolutions to 3D 3×3×1 convolutions.Source: Figure 3 · p. 5, §3 · p. 6
  • The study follows direct 3D segmentation with a U-shape FCN rather than classification-derived activation maps.The model segments ground-glass-like opaque regions from 3D chest CT.Source: §3 · p. 5
  • Each voxel is predicted as foreground or background, with a two-channel probability map produced after soft-max activation.Training uses Adam to minimize soft Dice loss; inference uses 224 × 224 × 32 crops with sliding steps of 16 along three axes.Source: §3 · p. 6

4. Federated Semi-Supervised Learning

The framework combines federated learning with consistency-based self-supervision so clients can train on private, heterogeneous CT data, including data without annotations. It uses weighted client updates and applies implementation choices to support training across supervised and unsupervised clients.

4.1. Federated Learning

  • The server aggregates client updates by weighted summation, while clients keep their CT data local and receive the updated model weights.Client weights account for iteration counts and user-assigned weights; the framework requires at least one labeled client to train the segmentation model.Source: §4.1 · pp. 6–7, §4.2 · p. 9

4.2. Federated Semi-Supervised Learning on COVID region segmentation

  • Unlabeled clients train from a global model by matching thresholded pseudo-labels to predictions on augmented CT images.The pseudo-label is generated by hard thresholding the current model prediction at 0.5; a confidence threshold τ ∈ [0.5, 1] can restrict regions used for the loss.Source: §4.2 · pp. 8–9
  • Federated semi-supervision faces weight divergence, delayed communication, client-weight and learning-rate choices, and non-iid data distributions.The text notes that patient populations, scanning protocols, scanners, and disease stages can differ across clients.Source: §4.2 · p. 9

4.3. Implementation Details

  • For one supervised and one unsupervised client, the supervised client is pretrained for 500 epochs before federation, followed by 500 federated epochs.With two supervised clients, training starts from scratch for 1,000 epochs; unsupervised clients use a 5e−6 learning rate and weights no larger than supervised clients to avoid slow or unstable convergence.Source: §4.3 · pp. 9–10

5. Experiment and Results

Federated semi-supervised learning was evaluated on three heterogeneous COVID-19 CT cohorts with uneven annotation availability. Results indicate that its performance depends on client data and design choices, including annotation quality, learning rates, and aggregation strategy.

5.1. Data and Expert Annotation

  • The study used COVID-19 CT scans from China, Japan, and Italy, with expert annotations on 671, 88, and 186 scans, respectively.The three cohorts differed in acquisition settings, and Image 3 annotations involved radiologists from different countries.Source: §5.1 · p. 10

5.2. Analysis

  • Single-site models showed limited cross-domain generalizability: models trained on Images 2 or 3 had a 10% Dice drop on Image 1 versus the Image 1-trained model.Image 3 may lower performance when added to training, potentially because of annotation-protocol bias, while an unsupervised Image 3 client could improve Image 2 generalizability; centralized two-GPU training performed much better than other settings.Source: §5.2 · pp. 11, 13
  • Approximately 1% Dice’s score drop versus supervised FL; federated semi-supervised learning generally outperformed models trained independently on Images 1, 2, or 3.On Image 2 and Image 3 testing results, unlabeled data from other clients were valuable for training a generalizable model; Fig. 5 shows better-shaped masks with fewer false positives.Source: Table 1 · p. 11, Figure 5 · p. 12, §5.2 · p. 12

5.3. Ablation Studies

  • Hard Dice generally edged other tested consistency losses, while foreground-only loss did not consistently improve performance; stronger Gaussian noise and larger augmentation differences improved generalizability.Increasing Cutout regions significantly reduced performance on Image 3, and the best augmentation strategy may depend on the task and data.Source: §5.3 · p. 13
  • Unsupervised-client aggregation weight had a non-linear relationship with performance, while sharing fewer layers performed better and encoder-only sharing generalized across seen and unseen datasets.The authors suggest unsupervised updates contain both Image 2 characteristics and noise; partial sharing may also offer a safer privacy-preserving option.Source: §5.3 · pp. 13–14
  • Lower unsupervised-client learning rates generally benefited all clients, whereas more frequent aggregation improved unsupervised-client performance but slightly worsened supervised-client performance.The aggregation-frequency trade-off with communication cost needs tuning for the application; ablation configurations were each trained once rather than repeated five times.Source: §5.3 · pp. 13, 15, 16

6. Conclusions and Discussion

The federated semi-supervised framework uses information from unlabeled clients while preserving patient privacy, and the authors report improved generalizability across client and unseen domains. Its scope remains segmentation of COVID-19-associated abnormal regions, not distinguishing them from other abnormalities.

6. Conclusions and Discussion

  • Joint training with supervised and unlabeled clients improved generalizability on each client’s database and on unseen domains, while preserving patient privacy.The framework also uses information from clients with only unlabeled data.Source: §6 · p. 16
  • The target is abnormal regions in COVID-19 cases; the model is not trained to distinguish them from other abnormalities such as pneumonia or cancer.The discussion describes additional classification as a possible pipeline stage.Source: §6 · p. 16
  • On LIDC, the model detected solid and ground-glass nodules and other abnormal patterns; the authors attribute nodule detections to their similar CT appearance.Figure 8 illustrates the listed abnormality categories.Source: §6 · p. 17, Figure 8 · p. 17
  • The framework was validated on a multinational cohort with population, equipment, and demographic variation, and the study included comprehensive ablation studies.Source: §6 · p. 17
  • The authors identify unresolved questions about domain gaps between supervised and unsupervised clients and aggregating client contributions by data quality rather than quantity.Source: §6 · p. 17
  • Figure 9 shows predicted COVID-19 regions with lungs and airways in 3D and in axial, sagittal, and coronal views; slice thickness is 5 mm.The passage contrasts this with 0.8 mm in-plane resolution and notes that 5 mm slices are typical of most images in the work.Source: Figure 9 · p. 17

Related papers

Loading 2011.11750v1…