COVID-Net USPro: An Open-Source Explainable Few-Shot Deep Prototypical Network to Monitor and Detect COVID-19 Infection from Point-of-Care Ultrasound Images

Technical Summary of COVID-NET USPRO

Problem Statement
The rapid transmission of SARS-CoV-2 necessitates effective screening methods to alleviate the burden on healthcare systems. While medical imaging, particularly Point-of-Care Ultrasound (POCUS), offers a portable, low-cost, and radiation-free alternative to CT and X-ray for detecting COVID-19, the application of deep learning to this modality faces a critical bottleneck: the scarcity of well-annotated data for novel diseases. Standard deep learning architectures typically require massive datasets, which are unavailable during the early stages of pandemics. Furthermore, existing solutions often lack interpretability and clinician validation, raising concerns about their adoption in real-world clinical settings.

Methodology
To address these challenges, the authors propose COVID-Net USPro, an open-source, explainable, few-shot deep prototypical network designed to detect COVID-19 from lung ultrasound (LUS) images using minimal training data.

• Data Source: The study utilizes the COVIDx-US dataset v1.4, comprising 29,651 processed images from 242 videos. The dataset includes four classes: COVID-19 positive, non-COVID-19 infection, other lung conditions, and normal controls. To ensure consistency and avoid data leakage, only images from convex probes were used (linear probe data was excluded due to field-of-view differences), and the data was split into 90% training and 10% testing sets. The dataset was augmented via 90°, 180°, and 270° rotations.
• Network Architecture: COVID-Net USPro is a Prototypical Network (PN) operating in an episodic learning setting. It functions by mapping input images into an embedding space where samples of the same class cluster around a single prototype.
  • Encoder: The network employs transfer learning using pre-trained models (ResNet18 and ResNet50) initialized on ImageNet. The authors experimented with varying the number of trainable layers (e.g., only the final connected layer vs. the last 3-4 convolutional layers plus the final layer).
  • Learning Mechanism: In each training episode, a "support set" (labeled examples) and a "query set" (unlabeled examples) are sampled. Class prototypes are computed as the mean of the support set embeddings. Query images are classified by calculating the squared Euclidean distance to these prototypes and applying a SoftMax function.
  • Training Configuration: The model was trained using an Adam optimizer with early stopping. The study evaluated performance across different "shot" scenarios (5, 10, 20, up to 200 shots) and classification complexities (2-way, 3-way, and 4-way).
• Explainability and Validation: To ensure the network relies on clinically relevant features rather than artifacts, the authors integrated two Explainable AI (XAI) tools: Grad-CAM and GSInquire. These tools generate visual heatmaps highlighting the regions influencing the classification decision. A contributing clinician (an Assistant Professor in Emergency Medicine) reviewed these visualizations to verify that the network focused on pathological indicators (e.g., B-lines, lung consolidation) rather than irrelevant structures (e.g., ribs).

Key Contributions

1. High-Performance Few-Shot Learning: The authors present a network capable of achieving high precision and recall with extremely limited data. When trained with only 5 shots per class, COVID-Net USPro achieved 99.65% overall accuracy, 99.7% recall, and 99.67% precision for COVID-19 positive cases in a 2-way classification task.
2. Explainability and Clinical Validation: Unlike many prior works, this study rigorously validates the network's decision-making process. Through Grad-CAM and GSInquire analysis, combined with clinician review, the authors demonstrated that the network can be guided to focus on disease-specific artifacts (such as B-lines and consolidation) rather than spurious correlations.
3. Open-Source Initiative: The model, code, and analysis pipeline are released as open-source to promote reproducibility and accessibility in healthcare AI.

Results

• Performance Scaling: Performance metrics improved as the number of training shots increased, plateauing after 75 shots. The ResNet50L4 configuration (pre-trained ResNet50 with the last 3 convolutional layers and final layer re-trained) consistently outperformed other encoders across all classification scenarios (2-way, 3-way, and 4-way).
• Classification Complexity: While increasing the number of classes (from 2-way to 4-way) initially reduced performance, the gap narrowed significantly with higher shot counts (e.g., 100 shots), where the model achieved near-perfect scores (1.0000) in 3-way and 4-way scenarios.
• Explainability Improvements: Initial visualizations showed instances where the model incorrectly focused on non-pathological features (e.g., ribs). By implementing two strategies—filtering images based on the Lung Ultrasound Score (LUSS) to ensure high-quality data and cropping the images to remove soft tissue above the pleura—the network's focus shifted to clinically relevant regions (the lung parenchyma below the pleura). The clinician confirmed that the refined model correctly identified B-lines and consolidation patterns.

Significance and Claims
The paper claims that COVID-Net USPro offers a viable solution for screening COVID-19 in resource-limited settings where large annotated datasets are unavailable. By leveraging few-shot learning, the network reduces the computational and data requirements typically associated with deep learning. The authors emphasize that the combination of high accuracy with explainable decision-making and clinician validation is crucial for the trust and adoption of AI in medical diagnostics. The work positions itself not just as a diagnostic tool for COVID-19, but as a framework adaptable to other novel diseases and pandemics where data is scarce. The authors modestly note that future work could involve more sophisticated image segmentation, inclusion of linear probe data, and the use of systematically collected, high-quality standardized datasets to further refine performance.