PaperlayerView exact revision on arXiv

Source-linked AI summary

Iteratively Pruned Deep Learning Ensembles for COVID-19 Detection in Chest X-rays

Sivaramakrishnan Rajaraman, Jen Siegelman, Philip O. Alderson, Lucas S. Folio, Les R. Folio, Sameer K. Antani

arXiv:2004.08379v3eess.IVcs.CVcs.LGstat.ML

TL;DR

COVID-19 screening needs accessible, high-throughput tools where CT availability or radiological expertise is limited. This study transfers CXR-specific features, prunes deep-learning models, and ensembles them, with weighted averaging outperforming other ensemble methods.

  • Problem

    Accessible automated CXR tools could support COVID-19 triage where CT availability or radiological expertise is insufficient.

  • Method

    The study transfers modality-specific CXR features, iteratively prunes task-specific deep-learning models, and combines pruned models using ensemble strategies.

  • Results

    Weighted averaging of the pruned models outperformed the other evaluated ensemble methods, while pruning improved prediction performance and reduced trainable parameters.

  • Takeaways & Limitations

    The combined approach supports a smaller, faster model with potential deployment on browsers and mobile devices for CXR-based respiratory disease screening.

Abstract

from arXiv · show

We demonstrate use of iteratively pruned deep learning model ensembles for detecting pulmonary manifestation of COVID-19 with chest X-rays. This disease is caused by the novel Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) virus, also known as the novel Coronavirus (2019-nCoV). A custom convolutional neural network and a selection of ImageNet pretrained models are trained and evaluated at patient-level on publicly available CXR collections to learn modality-specific feature representations. The learned knowledge is transferred and fine-tuned to improve performance and generalization in the related task of classifying CXRs as normal, showing bacterial pneumonia, or COVID-19-viral abnormalities. The best performing models are iteratively pruned to reduce complexity and improve memory efficiency. The predictions of the best-performing pruned models are combined through different ensemble strategies to improve classification performance. Empirical evaluations demonstrate that the weighted average of the best-performing pruned models significantly improves performance resulting in an accuracy of 99.01% and area under the curve of 0.9972 in detecting COVID-19 findings on CXRs. The combined use of modality-specific knowledge transfer, iterative model pruning, and ensemble learning resulted in improved predictions. We expect that this model can be quickly adopted for COVID-19 screening using chest radiographs.

I. INTRODUCTION

COVID-19 screening with portable chest X-rays could help address diagnostic constraints where CT is less available or practical. The study combines modality-specific deep learning, pruning, and ensembles to distinguish normal, bacterial-pneumonia, and COVID-19-related CXRs.

  • RT-PCR is the diagnostic gold standard, but its sensitivity varies and availability may be limited in some regions.
  • Portable CXRs are less sensitive than CT but can reduce isolation, sanitation, and personnel-exposure burdens.
  • Automated CXR-based CADx could support high-throughput triage where CT access or radiological expertise is insufficient.
  • The study distinguishes CXRs showing COVID-19 pneumonia-related opacities from bacterial pneumonia and normal lungs using publicly available collections.
  • The approach combines modality-specific knowledge transfer, iterative pruning, and ensemble strategies to improve robustness, generalization, and inference capability.

II. PRIOR WORK

Prior work covers AI-based COVID-19 imaging, modality-specific transfer learning, neural-network pruning, and ensemble classification. These strands motivate adapting and combining compact CXR models for COVID-19 detection.

  • COVID-19 detection: Earlier studies used CT imaging and deep CNNs to distinguish COVID-19 manifestations or classify CTs as normal versus COVID-19 pneumonia.
  • Modality-specific knowledge transfer: Modality-specific transfer learning addresses medical-image appearance differences by retraining CNNs on large CXR collections before fine-tuning.
  • Model pruning: Model pruning reduces computational complexity and can support low-resource deployment without loss or with improvement in performance.
  • Ensemble classification: Ensemble learning combines predictions from multiple CNNs to alleviate sensitivity to initialization and statistical noise in training data.
  • Ensemble classification: Prior CXR studies applied CNN ensembles to tuberculosis and cardiomegaly detection.

III. MATERIALS AND METHODS

The retrospective analysis combines four publicly available CXR collections covering normal, bacterial-pneumonia, non-COVID viral, and COVID-19-related findings. Data were split at the patient level for training, testing, and validation.

  • A. DATA COLLECTION AND PREPROCESSING: The study uses four publicly available CXR collections in a retrospective analysis.
  • A. DATA COLLECTION AND PREPROCESSING: The pediatric collection contains AP CXRs from children aged 1–5 years with normal, bacterial-pneumonia, and non-COVID viral-pneumonia findings.
  • A. DATA COLLECTION AND PREPROCESSING: The RSNA collection includes normal CXRs and abnormal images with non-pneumonia or pneumonia-like opacities at 1024 × 1024 pixel resolution in DICOM format.
  • A. DATA COLLECTION AND PREPROCESSING: The Twitter COVID-19 and Montreal collections provide COVID-19-related CXRs, with the Montreal repository containing 179 such images as of April 7, 2020.
  • A. DATA COLLECTION AND PREPROCESSING: Patient-level splits allocated 90% of the collections to training and 10% to testing, with 10% of training data used for validation.

B. LUNG ROI SEGMENTATION

The method segments lung regions, trains custom and pretrained CNNs with CXR-specific transfer learning, then fine-tunes selected models for three-way classification. It also uses dropout, separable convolutions, and tuned optimization settings.

  • B. LUNG ROI SEGMENTATION: U-Net semantic segmentation isolates lung pixels, and the resulting masks define bounding-box lung crops resized to 256×256 pixels.
  • B. LUNG ROI SEGMENTATION: The U-Net encoder uses Gaussian dropout with an empirically selected dropout ratio of 0.2.
  • C. MODELS AND COMPUTATIONAL RESOURCES: The model evaluation includes a customized CNN and eight ImageNet-pretrained CNN architectures, including VGG-16, VGG-19, and Inception-V3.
  • C. MODELS AND COMPUTATIONAL RESOURCES: The custom CNN stacks strided separable convolutions, global average pooling, dropout, and a Softmax dense layer.
  • D. MODALITY-SPECIFIC TRANSFER LEARNING AND FINE-TUNING: Modality-specific transfer learning retrains models on RSNA normal-versus-abnormal CXRs before transferring and fine-tuning them on the related classification task.
  • D. MODALITY-SPECIFIC TRANSFER LEARNING AND FINE-TUNING: The pretrained models are truncated at their deepest convolutional layers, augmented with task-specific layers, and fine-tuned to classify normal, bacterial-pneumonia, or COVID-19 viral-pneumonia CXRs.

E. ITERATIVE MODEL PRUNING

Iterative pruning reduces convolutional model complexity by removing low-activation neurons and retraining after each pruning step. The procedure evaluates and stores progressively pruned models up to a maximum pruning percentage.

  • Pruning criterion: The method ranks convolutional neurons using average percentage of zeros (APoZ), measured on validation-dataset activations.APoZ is the percentage of zero neuron activations observed with the validation dataset.
  • Stopping rule: The process continues while the percent pruned remains at or below the maximum pruning percentage M.The pruning percentage increases incrementally by P after each iteration.
  • Pruning procedure: At each iteration, the algorithm removes a fixed percentage of filters with the highest APoZ from every convolutional layer.The algorithm first calculates filter counts, identifies high-APoZ filters, and deletes the specified percentage.
  • Pruning procedure: Each pruned network is retrained and evaluated on the training set B, with its best-pruned weights saved.Retraining follows every pruning step before the next incremental iteration.
  • Outputs: The algorithm returns M+1 pruned models, including the base model and the progressively pruned versions.The returned models provide candidates for later ensemble construction.

F. LEARNING ITERATIVELY PRUNED ENSEMBLES

The best-performing pruned models are combined into ensembles for CXR classification. Multiple strategies combine predictions to target normal, bacterial pneumonia, and COVID-19 viral pneumonia-related opacities.

  • Ensemble construction: The ensemble selects the best-performing pruned models to improve prediction performance and generalization over individual constituent models.Model selection occurs after iterative pruning and focuses on the strongest pruned candidates.
  • Classification task: The ensemble classifies CXRs as normal or showing bacterial or COVID-19 viral pneumonia-related opacities.The classification task includes three CXR outcome categories.
  • Combination strategies: The evaluated ensemble strategies are max voting, averaging, weighted averaging, and stacking.Stacking uses a neural network-based meta-learner.

G. VISUALIZATION STUDIES

The study visualizes the learned behavior of pruned deep-learning models using gradient-based class-activation analysis. Salient regions of interest are compared with consensus annotations from experienced radiologists.

  • Visualization method: Grad-CAM computes image-category scores from feature maps in the deepest convolutional layer of a trained model.It is presented as a gradient-based visualization method for medical visual recognition.
  • Visualization method: Grad-CAM globally pools backward-flowing gradients to estimate the importance of weights in the decision-making process.The procedure links gradient information to model decision relevance.
  • Model verification: The pruned models’ salient regions of interest are compared with consensus ground-truth annotations from experienced radiologists.This comparison is used to verify the learned behavior of the models.

H. STATISTICAL ANALYSES

The statistical analysis evaluates CNN skill using confidence intervals. The study uses 95% Clopper–Pearson exact intervals and computes them with StatsModels version 0.11.0.

  • Confidence-interval analysis: Confidence intervals are used to analyze the skill of CNN models at different stages of learning.The analysis examines model performance across learning stages rather than at only one stage.
  • Interpretation: A shorter confidence interval indicates a smaller margin of error and a relatively more precise estimate.Larger intervals allow more margin for error and indicate reduced precision.
  • Interval computation: The study computes 95% Clopper–Pearson exact intervals as two-sided intervals with individual coverage probabilities of (0.95)1/2.StatsModels version 0.11.0 is used to compute the confidence-interval measures.

IV. RESULTS AND DISCUSSION

The study optimizes custom and pretrained CNNs for modality-specific CXR representations, then evaluates the strongest transferred models for COVID-19-related classification tasks.

  • IV. RESULTS AND DISCUSSION: The study optimized parameters and hyperparameters for customized and pretrained CNNs using Talos and randomized grid search, respectively.The optimized values are reported in Table 2.
  • IV. RESULTS AND DISCUSSION: VGG-16, VGG-19, and Inception-V3 were more accurate than the other evaluated models and provided stronger AUC, F-score, and MCC performance.Their AUC estimates also had shorter confidence intervals than those of the other models.
  • IV. RESULTS AND DISCUSSION: The top three modality-specific models were transferred with modality-specific weights, truncated at their fully connected layers, and fitted with task-specific heads.These models were evaluated for binary normal-versus-COVID-19 classification and multiclass normal, bacterial-pneumonia, or COVID-19 classification.

KNOWLEDGE TRANSFER MODELS ON THE TARGET TASKS

On the target tasks, the transferred models achieved perfect binary accuracy, while multiclass performance was evaluated with individual, pruned, and ensemble models. Weighted averaging of pruned models delivered the strongest overall ensemble performance.

  • KNOWLEDGE TRANSFER MODELS ON THE TARGET TASKS: 100% accuracy was achieved by all models on the binary task, while VGG-16 used the fewest trainable parameters.For the multiclass task, Inception-V3 was more accurate and had a shorter AUC confidence interval than VGG-16 and VGG-19.
  • KNOWLEDGE TRANSFER MODELS ON THE TARGET TASKS: The multiclass task combined VGG-16, VGG-19, and Inception-V3 predictions using max voting, simple averaging, weighted averaging, and model stacking.Ensembling was not performed for the binary task because the individual models were 100% accurate.
  • KNOWLEDGE TRANSFER MODELS ON THE TARGET TASKS: 46.03% fewer parameters were used by the pruned VGG-16 model than by its unpruned counterpart.Pruned VGG-19 and Inception-V3 reduced trainable parameters by 16.13% and 36.1%, respectively, while improving accuracy, F-score, and MCC.
  • KNOWLEDGE TRANSFER MODELS ON THE TARGET TASKS: Weighted averaging of pruned-model predictions delivered superior performance across the reported evaluation aspects.The ensemble assigned weights of 0.5, 0.3, and 0.2 to Inception-V3, VGG-19, and VGG-16 predictions, respectively.
  • KNOWLEDGE TRANSFER MODELS ON THE TARGET TASKS: The weighted-average pruned ensemble produced the reported confusion matrix and ROC curves, with the ROC figure including micro-, macro-, and class-specific AUC views.Its AUC confidence interval had the shortest error margin among the ensemble methods, and it outperformed the alternatives on F-score and MCC.

V. CONCLUSION

The study combines CXR modality-specific knowledge transfer, iterative pruning, and ensemble learning for COVID-19 pneumonia detection. This combination improved prediction performance and deployment-related efficiency, while success remained dependent on dataset size and computational resources.

  • Ensemble deep learning applied to CXR findings supports detection of COVID-19 pneumonia alongside bacterial pneumonia and normal cases.
  • Modality-specific transfer learning improved model adaptation and generalization compared to ImageNet pretrained weights in a related CXR classification task.
  • Iterative pruning improved test-data prediction performance while significantly reducing trainable parameters, producing smaller and faster models.The authors connect this reduction to removing redundant, low-activation neurons and describe browser and mobile deployment as a use case.
  • Weighted averaging of pruned models outperformed the other evaluated ensemble methods.
  • Visualization studies found that pruned models precisely localized salient regions of interest used to categorize input CXRs.
  • The approach’s success is controlled by dataset size and variability, along with computational resources required for training and practical deployment.The authors identify the number and diversity of relevant viral-pneumonia CXRs, training time, and memory constraints as practical considerations.
Loading 2004.08379v3…