PaperlayerView exact revision on arXiv

Source-linked AI summary

Contrastive Cross-site Learning with Redesigned Net for COVID-19 CT Classification

Zhao Wang, Quande Liu, Qi Dou

arXiv:2009.07652v1eess.IVcs.CVcs.LG

TL;DR

COVID-19 CT diagnosis needs automated tools and robust learning from heterogeneous multi-site datasets. The paper redesigns COVID-Net, uses site-specific feature normalization and contrastive learning, and reports consistent improvements on two public datasets, including AUC gains of 12.16% and 14.23% over original COVID-Net.

  • Problem

    Automated COVID-19 CT analysis is needed to support diagnosis and reduce radiologists’ image-interpretation workload, while heterogeneous multi-site data complicate robust model development.

  • Method

    The framework redesigns COVID-Net for CT images, applies separate feature normalization for each site, and uses contrastive learning to promote domain-invariant semantic representations.

  • Results

    12.16% and 14.23% AUC improvements over original COVID-Net are reported on the two datasets, alongside consistent classification gains on both datasets.

  • Takeaways & Limitations

    Joint learning with separate normalization and contrastive training improves COVID-19 CT classification across heterogeneous datasets and shows promising lesion localization across sites.

Abstract

from arXiv · show

The pandemic of coronavirus disease 2019 (COVID-19) has lead to a global public health crisis spreading hundreds of countries. With the continuous growth of new infections, developing automated tools for COVID-19 identification with CT image is highly desired to assist the clinical diagnosis and reduce the tedious workload of image interpretation. To enlarge the datasets for developing machine learning methods, it is essentially helpful to aggregate the cases from different medical systems for learning robust and generalizable models. This paper proposes a novel joint learning framework to perform accurate COVID-19 identification by effectively learning with heterogeneous datasets with distribution discrepancy. We build a powerful backbone by redesigning the recently proposed COVID-Net in aspects of network architecture and learning strategy to improve the prediction accuracy and learning efficiency. On top of our improved backbone, we further explicitly tackle the cross-site domain shift by conducting separate feature normalization in latent space. Moreover, we propose to use a contrastive training objective to enhance the domain invariance of semantic embeddings for boosting the classification performance on each dataset. We develop and evaluate our method with two public large-scale COVID-19 diagnosis datasets made up of CT images. Extensive experiments show that our approach consistently improves the performances on both datasets, outperforming the original COVID-Net trained on each dataset by 12.16% and 14.23% in AUC respectively, also exceeding existing state-of-the-art multi-site learning methods.

I. INTRODUCTION

COVID-19 CT diagnosis needs automated tools and larger, more diverse datasets because clinical workloads and single-site data limitations impede robust model development. The paper proposes joint learning with redesigned COVID-Net, separate feature normalization, and contrastive representations for heterogeneous datasets.

  • Automated COVID-19 CT analysis is sought to support diagnosis, patient management, and radiologists facing overloaded services and limited trained personnel.
  • Aggregating CT data across hospitals is proposed to address insufficient single-site sample sizes and enable cross-site learning.
  • Separate feature normalization estimates site-specific statistics, while contrastive learning promotes category-sensitive, domain-invariant semantic features.
  • Experiments on two public datasets consistently improve classification performance on both datasets and outperform single-site training, straightforward joint learning, and existing multi-site methods.
  • The framework redesigns COVID-Net for improved computational efficiency and prediction accuracy on COVID-19 CT images.

II. RELATED WORKS

Prior COVID-19 image-classification research used established networks, specialized architectures, and multi-site datasets. These studies motivate combining data sources while addressing the heterogeneity that simple aggregation may overlook.

  • Earlier CT-based screening methods distinguished COVID-19 pneumonia from influenza-A pneumonia and healthy cases using ResNet18 with location attention.
  • Transfer-learning studies commonly used established architectures including VGG, ResNet, and DenseNet, while MobileNet-based work emphasized interpretability.
  • COVID-Net was tailored for COVID-19 recognition on chest X-rays, while CovidCTNet and Capsule Network approaches targeted CT diagnosis and small-scale datasets.
  • Multi-site research explored hypergraph models, very large datasets, and combined CT and chest X-ray sources for COVID-19 diagnosis and severity analysis.

III. METHODS

The proposed framework redesigns COVID-Net as a CT-classification backbone and combines site-specific normalization with contrastive learning. These components target computational efficiency, cross-site heterogeneity, and domain-invariant semantic embeddings.

  • The method redesigns COVID-Net and integrates joint learning, separate feature normalization, and contrastive loss for heterogeneous COVID-19 datasets.
  • Domain Invariance Enhancement: The framework uses separate normalization across sites and contrastive learning to improve classification by making latent semantic embeddings domain invariant.
  • COVID-Net Redesign: COVID-Net was originally designed for chest X-rays, whereas CT images contain clearer lesion patterns and richer information requiring adapted architecture and learning strategies.
  • COVID-Net Redesign: The redesign adds batch normalization selectively to reduce internal feature variance, improve feature discrimination, and accelerate convergence without unnecessarily increasing computation.

1) Network architecture redesign:

The architecture redesign adds selective normalization and global average pooling to stabilize CT feature learning and reduce classifier complexity. These choices address feature variance, training efficiency, and overfitting concerns.

  • Network architecture redesign: Batch normalization is added to the initial convolution and upper-branch convolutional layers rather than every layer.
  • Network architecture redesign: Batch normalization transforms channel-wise feature maps using normalized values with trainable affine parameters.
  • Network architecture redesign: During testing, batch normalization uses moving-average statistics collected during training to normalize features.
  • Network architecture redesign: Global average pooling produces compact high-level embeddings and reduces output dense-layer parameters by 12 times in this architecture.
  • Network architecture redesign: CT lesion appearance varies substantially with disease severity, making robust optimization across heterogeneous datasets more difficult.

2) Learning strategy redesign:

The framework adds a joint learning scheme to aggregate heterogeneous data sources, while redesigning COVID-Net and using domain-specific normalization to address cross-site statistical differences.

  • The joint learning scheme aggregates cases from different data sources to address insufficient COVID-19 samples from individual hospitals.
  • The redesigned COVID-Net backbone is used to improve computational efficiency and recognition performance before joint learning.
  • Domain-specific batch normalization addresses statistical differences across datasets that can make joint-model moving averages inaccurate during training.

1) Separate batch normalization at data heterogeneity:

The method targets heterogeneous-site representations by learning semantic embeddings that remain class-consistent across domains through contrastive learning on projected features.

  • The framework seeks semantic embeddings that cluster by class regardless of data source domain, so multi-site data can be projected into a harmonized feature space.
  • An embedding network projects pooled semantic features into a lower-dimensional space before pairwise similarity is computed.
  • Positive pairs contain samples from the same class, whereas negative pairs contain samples from different classes.
  • The contrastive objective promotes within-class cohesion and between-class separation, encouraging domain-invariant embeddings across the two sites.

C. Overall Training Objective and Technical Details

Training combines classification and contrastive objectives and evaluates the framework on two public CT datasets using cross-validation and five reported metrics.

  • The overall objective combines cross-entropy classification loss with contrastive loss to assess classification error and regularize the latent space.
  • Cross-entropy uses one-hot ground-truth labels and predicted probabilities, while the embedding network has fully connected outputs of 1024 and 128 dimensions.
  • The framework uses Adam optimization from scratch, with an initial learning rate of 1e-4, cosine annealing, temperature τ=0.05, and α=1.0.
  • Evaluation uses the SARS-CoV-2 and COVID-CT public datasets, with images resized to 224 × 224 and normalized to zero mean and unit variance.
  • Four-fold cross-validation reports mean and standard deviation across three independent runs for accuracy, F1 score, sensitivity, precision, and AUC.

B. Effectiveness of Network Redesign on COVID-Net

The redesigned COVID-Net backbone outperforms the original COVID-Net across the two sites and evaluation settings, with improvements across nearly all reported measures.

  • The redesigned model consistently improves all five evaluation metrics over the original COVID-Net in the Single setting on both sites.
  • In the Joint setting, the redesigned model also improves performance over the original backbone, except for a marginal precision improvement in Site B.
  • The redesigned architecture and learning strategy show an advantage in ROC curves across both Single and Joint settings.
  • The comparisons use Single models trained per site and Joint models trained by naively aggregating the two datasets.

1) Comparison with baseline settings:

The proposed method improves joint COVID-19 CT classification over baseline and state-of-the-art multi-site approaches across both sites. Separate normalization and contrastive training support gains that are statistically significant and broadly consistent across evaluation metrics.

  • Baseline comparisons: 96.24% and 85.32% AUC are achieved on Site A and Site B after adding contrastive training to separate normalization.The contrastive objective further improves performance over SepNorm on both sites.
  • Baseline comparisons: Our method outperforms Single in 9 out of 10 metrics across the two sites.This supports using heterogeneous datasets jointly rather than relying only on separate single-site models.
  • Baseline comparisons: All paired t-tests report p-values below 0.05 against Joint, Single, and SepNorm on both sites.The reported improvements are statistically significant under the paired t-test analysis.
  • State-of-the-art comparisons: Figure 3 compares ROC curves across backbone designs, baseline settings, and other joint-learning methods for Site A and Site B.The panels use upper and lower plots for Site A and Site B, respectively.
  • State-of-the-art comparisons: Parallel Adapter improves AUC over Joint by 4.27% on Site A and 7.81% on Site B, while Series Adapter remains below Single.These comparisons indicate that domain-specific adapter parameters help address heterogeneous data, but performance gains can be imbalanced.
  • State-of-the-art comparisons: The method considerably outperforms all three state-of-the-art joint learning methods on both sites.Paired t-tests in Table III indicate statistically significant improvements over those comparison methods.

V. DISCUSSIONS

The discussion frames multi-site COVID-19 CT diagnosis as difficult because lesion appearance varies and pneumonia patterns overlap. Grad-CAM suggests useful lesion localization, while failure cases and limited site coverage constrain deployment.

  • Motivation and challenges: COVID-19 CT diagnosis is difficult because lesions vary substantially with severity and resemble other pneumonia types.These patterns motivate models trained on broad datasets that capture diverse samples and lesion distributions.
  • Method discussion: Redesigning COVID-Net adapts its architecture and learning strategy to capture complex CT lesion representations more effectively than its original X-ray-oriented setting.The redesign targets both recognition performance and learning efficiency.
  • Interpretability and failures: Grad-CAM consistently localizes suspicious regions across bilateral, peripheral ground-glass, and consolidative pulmonary opacity patterns, including mild lesions.The authors interpret this as promising interpretability and potential clinical relevance for image-based diagnosis.
  • Interpretability and failures: Failure cases arise from wrongly attended regions and difficulty distinguishing images with unobvious lesions.These examples identify concrete conditions under which the classification model remains unreliable.
  • Limitations: The method is limited to the two study sites and may not transfer directly to unseen sites.The authors identify wider cross-site deployment and generalization as unresolved challenges.
  • Limitations: The study did not pretrain the redesigned model on large-scale datasets such as ImageNet because of computational-resource and development-urgency constraints.Future work proposes combining the redesigned architecture with model transfer learning and extending evaluation to more sites.

VI. CONCLUSION

The paper develops a joint learning framework for COVID-19 CT diagnosis using heterogeneous datasets. It redesigns COVID-Net, separates site-specific normalization, and applies contrastive learning to improve domain-invariant features.

  • The framework redesigns COVID-Net in its architecture and learning strategy as a backbone for COVID-19 CT diagnosis.
  • Separate feature normalization explicitly mitigates inter-site data heterogeneity in joint learning.
  • Contrastive training enhances domain-invariant semantic features to improve identification performance on each dataset.
Loading 2009.07652v1…