Introduction
Accurately assessing the severity of COVID-19 is paramount for clinical decision-making and hospital resource management. While deep learning models have shown promise in analyzing CT scans, their real-world deployment is often hindered by variability across different clinical sites. This paper introduces FLARE (Fair cross-domain adaptation with LAtent REpresentations), a framework designed to provide robust severity assessment across diverse hospital environments by overcoming the dual hurdles of domain shift and class imbalance.
Challenges in Clinical Deployment
Two primary obstacles prevent the seamless transfer of diagnostic models between hospitals:
- Domain Shift: Differences in CT scanner manufacturers (e.g., Siemens vs. GE) and imaging protocols lead to variations in image contrast and noise levels. A model trained on one site's data may perform poorly on another's due to these subtle visual discrepancies.
- Class Imbalance: Clinical datasets typically contain a much higher proportion of mild cases compared to severe ones. Standard models often become biased toward the majority class, leading to poor sensitivity for identifying critical severe cases.
The FLARE Methodology
FLARE addresses these challenges through a unified approach that combines intelligent sampling with advanced representation learning.
Stochastic Class-balanced Boosting Sampling (SCBS)
To mitigate class imbalance, FLARE employs SCBS. Unlike static oversampling, SCBS dynamically adjusts the sampling probability during training. It prioritizes not only the minority "severe" class but also "hard" samples that the model currently finds difficult to classify, ensuring the model learns the subtle boundaries between severity levels.
Latent Representation Learning
The core of FLARE is a feature extractor that maps complex CT features into a latent space designed with three key properties:
- Domain-transferability: A domain translator and prototype triplet loss align data from different sites around shared "class centers" (prototypes), reducing the impact of domain shift.
- Discriminative Power: Conditional Maximum Mean Discrepancy (CMMD) loss ensures that "mild" and "severe" features remain well-separated in the latent space.
- Completeness: A multi-view reconstruction loss forces the latent features to retain sufficient information to reconstruct the original input, preventing the loss of vital clinical indicators during the translation process.
Experimental Evaluation
The researchers validated FLARE using a comprehensive dataset comprising 5,358 CT images from three Chinese hospitals and 1,110 scans from the public MosMedData dataset in Russia. Rather than using raw pixels, the model utilizes 237-dimensional handcrafted features—including texture, shape, and volume metrics—that align with expert radiological knowledge.
Key Results
FLARE consistently outperformed state-of-the-art domain adaptation methods such as DAN, DANN, and CDAN across multiple transfer tasks:
- Superior Performance: FLARE achieved G-mean scores (a balance of sensitivity and specificity) exceeding 86% on challenging cross-site tasks, significantly outperforming standard supervised approaches.
- Data Efficiency: The framework remains effective even when the target site has very limited labeled data, successfully "borrowing" knowledge from other sites.
- Scalability: An extended version, M-FLARE, demonstrated that performance improves as more source sites are integrated, effectively aggregating multi-site knowledge while ignoring the noise of different scanner types.
Conclusion
FLARE provides a robust solution for the cross-site assessment of COVID-19 severity. By simultaneously addressing domain shift and class imbalance through latent representation learning and adaptive sampling, it offers a practical path toward deploying reliable AI diagnostic tools in diverse clinical settings. Future work will explore end-to-end integration with automated segmentation and the analysis of longitudinal disease progression.
Summary updated