Abstract
Automatic severity assessment and progression prediction can facilitate admission, triage, and referral of COVID-19 patients. This study explores the potential of lung lesion features in COVID-19 management using the LesionEncoder framework. The framework consists of a U-Net module for lesion detection and feature extraction, and a recurrent neural network (RNN) for sequence-based classification. Results show high performance in severity assessment (AUC 0.903) and superior performance in progression prediction (AUC 0.736) compared to volumetric baselines.
Introduction
The COVID-19 pandemic placed unprecedented pressure on healthcare systems, necessitating rapid and accurate patient triage. Chest computed tomography (CT) is a vital tool for diagnosing and monitoring the disease. However, manual interpretation by radiologists is labor-intensive and subject to inter-observer variability. While many automated systems have been developed for severity assessment, predicting disease progression—identifying which mild cases will become severe—remains a significant challenge. Existing models often rely on lesion volume, which may not capture the subtle textural changes indicative of future clinical decline.
Methods
The researchers developed LesionEncoder, a deep learning framework designed to analyze the full 3D volume of a chest CT scan. The architecture follows a two-stage process:
- Lesion Encoding: A U-Net architecture with an EfficientNetB7 backbone is employed to segment lesions, including ground-glass opacities and consolidation. Beyond simple segmentation, the encoder extracts a 2,560-dimensional feature vector for each axial slice, capturing complex textural and morphological information.
- Sequence Modeling: To integrate information across the entire lung volume, a Bidirectional Long Short-Term Memory (BiLSTM) network is used. This RNN treats the sequence of slice-level feature vectors as input, allowing the model to learn the global context of the infection.
To manage computational complexity, Independent Component Analysis (ICA) was used to reduce the feature dimensionality before sequence classification. The model was trained on 433 CT scans from Huang Shi Central Hospital and independently tested on 206 scans from Xiang Yang Central Hospital.
Results
The LesionEncoder framework was evaluated on two distinct tasks: current severity assessment and future progression prediction.
Severity Assessment
The model successfully distinguished between mild and severe cases with an accuracy of 0.940 and an AUC of 0.903. It demonstrated high specificity (0.952), effectively identifying non-severe patients.
Progression Prediction
The most significant finding was the model's ability to predict "converters"—patients who were initially mild but progressed to severe illness within one week. LesionEncoder achieved an AUC of 0.736, significantly outperforming a baseline volumetric model (AUC 0.510), which performed no better than chance.
The study highlights that while lesion volume correlates with current severity, it is a poor predictor of future progression. The high-dimensional features extracted by LesionEncoder provide the necessary prognostic information that simple volume metrics lack.
Discussion and Conclusion
The LesionEncoder framework demonstrates strong potential for clinical application in COVID-19 management, particularly in identifying high-risk patients early. By leveraging sequence modeling and detailed lesion features, the system overcomes the limitations of traditional volumetric analysis.
Despite its success, the study notes limitations, including a small sample size of "converter" cases and the exclusion of non-imaging clinical data (e.g., age, lab results). Future work will focus on multi-modal data integration and validating the framework on larger, more diverse international datasets. This approach also holds promise for other lesion-focused medical image analysis tasks beyond COVID-19.
Summary updated