Source-linked AI summary
Diagnosis of Coronavirus Disease 2019 (COVID-19) with Structured Latent Multi-View Representation Learning
Hengyuan Kang, Liming Xia, Fuhua Yan, Zhibin Wan, Feng Shi, Huan Yuan, Huiting Jiang, Dijia Wu, He Sui, Changqing Zhang, Dinggang Shen
TL;DR
COVID-19 versus CAP diagnosis from CT images requires integrating heterogeneous features while addressing limited data and weaknesses in existing multi-view representations. The paper learns a structured latent representation that encodes complementary information and class structure, then uses it for classification. The method achieves the best reported performance among comparisons and remains stable as training-data proportions vary.
Problem
Existing methods may use single feature types or fail to ensure complete, separable multi-view representations, while deep models require large-scale data and may not exploit expert prior.
Method
The method learns a structured latent space from multiple CT feature types, using neural networks for feature integration and a regressor to map new subjects into that space before classification.
Results
The proposed method achieves the best comparison performance, reaching 95.50% accuracy, and shows stable results as the training-data proportion varies.
Takeaways & Limitations
The learned latent representation can be used with different classifiers and improves diagnosis compared with original features.
Abstract
from arXiv · showhide
Recently, the outbreak of Coronavirus Disease 2019 (COVID-19) has spread rapidly across the world. Due to the large number of affected patients and heavy labor for doctors, computer-aided diagnosis with machine learning algorithm is urgently needed, and could largely reduce the efforts of clinicians and accelerate the diagnosis process. Chest computed tomography (CT) has been recognized as an informative tool for diagnosis of the disease. In this study, we propose to conduct the diagnosis of COVID-19 with a series of features extracted from CT images. To fully explore multiple features describing CT images from different views, a unified latent representation is learned which can completely encode information from different aspects of features and is endowed with promising class structure for separability. Specifically, the completeness is guaranteed with a group of backward neural networks (each for one type of features), while by using class labels the representation is enforced to be compact within COVID-19/community-acquired pneumonia (CAP) and also a large margin is guaranteed between different types of pneumonia. In this way, our model can well avoid overfitting compared to the case of directly projecting highdimensional features into classes. Extensive experimental results show that the proposed method outperforms all comparison methods, and rather stable performances are observed when varying the numbers of training data.
I. INTRODUCTION
The paper targets COVID-19 versus CAP diagnosis from CT-derived features, addressing limitations in single-view learning and direct high-dimensional projection. It proposes structured multi-view latent representation learning to integrate complementary information and improve diagnostic performance.
- Research gap: Existing approaches may use only one feature type, while multi-view methods may lack information completeness and separable class structure, harming testing performance.The paper links these limitations to overfitting training data and poorer generalization.
- Approach: The proposed framework maps samples into a latent space that encodes complementary feature information and reflects class distribution instead of directly classifying original features.A projection model maps new subjects into the learned latent representation, followed by classification in that space.
- Contributions: Multi-view representation learning explores complementarity among different CT feature types for COVID-19 and CAP classification.The authors present this as a full diagnostic pipeline rather than direct projection into the class space.
- Contributions: The learned latent representation can be used with different classifiers and performs better than original features across the baseline classifiers tested.The representation is described as compact, structured, and more resistant to overfitting high-dimensional data.
- Results: 95.50% accuracy is reported for the proposed method, with improvements over comparison methods in accuracy, sensitivity, and specificity.The reported improvement in diagnosis metrics is based on extensive CT-image experiments.
II. MATERIAL
The study uses CT images from 2,522 cases of COVID-19 and CAP, with standardized thin-section scanning and 189 extracted features grouped by feature type.
- 2,522 CT images comprise 1,495 COVID-19 cases and 1,027 community-acquired pneumonia (CAP) cases.Patients range from 12 to 98 years old and were drawn from several Chinese hospitals.
- CT scans used thin sections and a protocol including 120 kV, 0.625–2 mm reconstructed thickness, and breath-hold at full inspiration.
- V-Net preprocessing extracted lung structures and infected lesions before 189-dimensional features were computed from each CT image.
- Volume and surface features characterize infected-area extent, infection and lung-boundary surfaces, and distances between them.
III. METHOD
The proposed diagnosis framework integrates heterogeneous CT feature types through latent representation learning, then maps subjects into that space for classification of COVID-19 and CAP.
- The pipeline uses multi-view learning because heterogeneous CT features provide complementary diagnostic information and cannot be effectively concatenated directly.
- Training has three components: Complete and Structured Representation Learning, a latent-representation regressor, and a latent-representation-based classifier.
- The framework learns latent representations that encode information from different feature types while reflecting class structure for COVID-19 and CAP.
- During testing, original features are transformed into latent representations before the classifier produces the final diagnosis.
A. Step-1: Complete and Structured Representation Learning
Step 1 learns low-dimensional latent representations that reconstruct multiple feature types and organize samples according to their COVID-19 or CAP labels.
- The training set consists of multi-view samples X_n with corresponding binary class labels y_n for COVID-19 or CAP.
- The learned representations are intended to be compact within COVID-19 and CAP groups while maintaining a clear margin between them.
- Each latent representation is trained to encode heterogeneous feature information by reconstructing every feature type through a separate mapping.
- Minimizing the reconstruction objective encourages the latent representation h_n to contain information from all feature types.
2) Structure for latent representation:
The structured-representation loss organizes latent features by pneumonia class, promoting within-class compactness and a margin between COVID-19 and CAP. The overall objective jointly considers informativeness and separability.
- The section specifies a loss for making learned latent representations well structured with respect to the two pneumonia diseases.
- The structured-representation loss enforces compactness within each pneumonia class and a margin between COVID-19 and CAP.
- The complete objective jointly considers informativeness and separability through the proposed optimization function.
- λ > 0 balances completeness against class-label information in the objective.
B. Step-2: Learning Projection from Original Features to Latent Representation
Step 2 trains a latent-representation regressor that maps original CT features to the learned latent space. The mapping is trained against the corresponding latent representations using mean squared error.
- The latent-representation regressor maps a subject’s original features into the corresponding low-dimensional latent representation.
- The regressor uses fully connected neural networks to learn the mapping Γ(·) from original features to latent representations.
- Training minimizes mean squared error between regressor outputs and their corresponding learned latent representations.
- Given multiple types of original CT-image features, the corresponding latent representation can be calculated after projection.
C. Step-3: Latent-Representation-Based Classifier
Step 3 trains a classifier on the projected latent representation to distinguish COVID-19 from CAP. The classifier is optimized with cross-entropy loss.
- The latent-representation-based classifier diagnoses subjects as having COVID-19 or community-acquired pneumonia.
- The classifier is a neural network with three fully connected layers that takes the projected latent representation as input.
- The classification task uses the widely used cross-entropy loss as its optimization objective.
D. Testing Stage
During testing, original feature sets are transformed into latent representations and then classified. Experiments use a 70%/30% train–test split, five-fold training cross-validation, and feature preprocessing comparisons.
- D. Testing Stage: Testing first transforms subjects’ different feature types into latent representations, then obtains diagnosis results with the latent-representation-based classifier.
- D. Testing Stage: The dataset is randomly divided into 70% training data and 30% testing data.
- D. Testing Stage: Five-fold cross-validation on the training data tunes λ over {0.1, 1, 10, 100}, with λ fixed at 100 in the experiments.
- D. Testing Stage: Preprocessing is necessary because the original CT-derived features have rather different scales.
- D. Testing Stage: Standardization and normalization improve accuracy by 1.32% ∼25.69% across classification models compared with original features.
1) Discrimination power of these different types of features:
The experiments show substantial discrimination differences among CT feature types, with radiomic gray and texture features outperforming number and intensity features. The learned latent representation is better structured than original features and improves diagnosis across classifiers and metrics.
- Feature discrimination: Radiomic gray and texture features achieve clearly better diagnostic performance than number and intensity features.The paper attributes the difference to stronger manifestations between COVID-19 and other pneumonias in gray and texture features.
- Latent representation: Original concatenated features are not well structured, whereas learned latent representations reveal the underlying class structure.This comparison is observed in the training and testing visualizations.
- Classifier evaluation: Both conventional models and neural networks significantly improve on all three metrics when using the learned latent representation.Gaussian-Naive-Bayes improves accuracy by 4.72% and sensitivity by 3.44%, while Logistic-Regression improves specificity by 3.13%.
- Comparison with baselines: The proposed method reaches 95.50% accuracy and improves accuracy over baselines by 6.1% ∼19.9%.It also improves sensitivity and specificity by 4.61% ∼21.22% compared with comparison methods.
- Classifier evaluation: Using the latent representation yields consistently better performance than original features across all evaluated classifiers and metrics.The result holds for both conventional learning models and neural networks.
4) Stability of proposed method:
The proposed method improves as the training set grows, then becomes relatively stable once more than 40% of data are used. At 60% training data it achieves the best results on all three metrics, while the worst result is only about 1% lower.
- Training-data stability: Performance improves as the training-data ratio increases from 2% to 80%, with stability emerging above 40%.The testing set remains fixed in each experiment, and the authors describe the pattern as diminishing returns.
- Training-data stability: At 60% training data, the model achieves the best results on all three metrics.The reported worst performance is only about 1% lower than the best.
- Overall performance: The conclusion reports diagnosis performance of 95.5% accuracy, 96.6% sensitivity, and 93.2% specificity.The study frames the pipeline as leveraging multiple CT feature types for automatic COVID-19 diagnosis.