Source-linked AI summary
Adaptive Feature Selection Guided Deep Forest for COVID-19 Classification with Chest CT
Liang Sun, Zhanhao Mo, Fuhua Yan, Liming Xia, Fei Shan, Zhongxiang Ding, Wei Shao, Feng Shi, Huan Yuan, Huiting Jiang, Dijia Wu, Ying Wei, Yaozong Gao, Wanchun Gao, He Sui, Daoqiang Zhang, Dinggang Shen
TL;DR
COVID-19 classification with chest CT could assist diagnosis amid clinician and radiologist shortages, but existing handcrafted and neural-network approaches have limitations on small medical datasets. The proposed AFS-DF combines deep-forest representations with adaptive feature selection and achieves 91.79% ACC for COVID-19 versus CAP classification.
Problem
Chest CT can assist COVID-19 diagnosis, while existing handcrafted features and neural-network representations have limitations for classification, especially with small medical image datasets.
Method
AFS-DF extracts location-specific CT features, learns high-level representations with a deep forest, and adaptively selects discriminative features to reduce redundancy.
Results
91.79% ACC was achieved for COVID-19 versus CAP classification, exceeding LR, SVM, RF, and NN baselines.
Takeaways & Limitations
AFS-DF achieved superior COVID-19 classification performance compared with the evaluated machine-learning comparison methods.
Takeaways & Limitations
The method was validated only on COVID-19 versus CAP classification and currently uses handcrafted features based on prior knowledge.
Abstract
from arXiv · showhide
Chest computed tomography (CT) becomes an effective tool to assist the diagnosis of coronavirus disease-19 (COVID-19). Due to the outbreak of COVID-19 worldwide, using the computed-aided diagnosis technique for COVID-19 classification based on CT images could largely alleviate the burden of clinicians. In this paper, we propose an Adaptive Feature Selection guided Deep Forest (AFS-DF) for COVID-19 classification based on chest CT images. Specifically, we first extract location-specific features from CT images. Then, in order to capture the high-level representation of these features with the relatively small-scale data, we leverage a deep forest model to learn high-level representation of the features. Moreover, we propose a feature selection method based on the trained deep forest model to reduce the redundancy of features, where the feature selection could be adaptively incorporated with the COVID-19 classification model. We evaluated our proposed AFS-DF on COVID-19 dataset with 1495 patients of COVID-19 and 1027 patients of community acquired pneumonia (CAP). The accuracy (ACC), sensitivity (SEN), specificity (SPE) and AUC achieved by our method are 91.79%, 93.05%, 89.95% and 96.35%, respectively. Experimental results on the COVID-19 dataset suggest that the proposed AFS-DF achieves superior performance in COVID-19 vs. CAP classification, compared with 4 widely used machine learning methods.
I. INTRODUCTION
The paper addresses the need for automated COVID-19 classification from chest CT when clinicians and radiologists are scarce. It proposes AFS-DF to learn high-level representations from location-specific features while adaptively removing redundant features.
- COVID-19 increases demand for automated chest-CT classification tools because clinicians and radiologists are in short supply.
- Existing approaches rely mainly on handcrafted features or deep neural-network representations, each with limitations for high-level information use or small medical datasets.
- AFS-DF uses a deep forest to learn latent high-level representations from location-specific chest-CT features with relatively small training data.
- Adaptive task-driven feature selection discards low-importance features based on trained forests, integrating selection with classifier training.
- The evaluation uses 1495 COVID-19 patients and 1027 CAP patients, and reports superior classification performance against comparison methods.
II. MATERIALS AND METHOD
The study evaluates chest-CT classification using a retrospective dataset comprising COVID-19 and CAP cases acquired across multiple hospitals and scanner systems.
- The dataset contains 1495 confirmed COVID-19 cases and 1027 CAP cases collected from five hospitals.
- All patients underwent thin-section chest CT scans using scanners from UIH, GE, Toshiba, Siemens, and Hitachi.
- The CT protocol used 120 kV, reconstructed slice thicknesses from 0.625 to 2 mm, and breath-holding at full inspiration.
- The study images were de-identified, and institutional review approval was obtained with informed consent waived because of the retrospective design.
B. Feature Extraction
The feature-extraction procedure represents chest CT scans using segmented infection locations, lesion distributions, intensity and texture patterns, surface-area information, and demographic variables.
- Chest CT images are automatically segmented into infected lung regions and bilateral lung fields using VB-Net.
- Volume features quantify infected-region burden across whole lungs, lobes, pulmonary segments, and left-right differences.
- Infected-lesion-number features capture multifocal involvement across bilateral lungs, lobes, and pulmonary segments.
- Histogram, intensity, texture, and surface-area features characterize image patterns within infected regions.
- Age and gender are included, producing a 239-dimensional location-specific feature representation.
C. Adaptive Feature Selection Guided Deep Forest
AFS-DF cascades random-forest layers that augment inputs with class probabilities and adaptively retain features judged important by the trained forests.
- Each deep-forest layer contains N independent random forests and an adaptive feature-selection unit.
- Each random forest outputs COVID-19 and CAP probability distributions, which are concatenated with the input feature vector.
- Feature importance is calculated for each input feature across the trained random forests before low-importance features are discarded.
- The selected feature vector becomes the next layer’s input, allowing feature selection and classifier training to operate in one adaptive framework.
- Cascaded layers learn a deep discriminative feature representation for COVID-19 classification.
D. Implementation
The implementation uses an ensemble of forests within AFS-DF to classify COVID-19 versus CAP from location-specific CT features. Training uses five-fold subdivision, while testing ensembles the final-layer forest probabilities with a MAP decision.
- D. Implementation: AFS-DF combines XGBoost, random forest, and extremely randomized tree ensembles, using 20, 20, 20, and 50 trees, respectively.The feature discard ratio is empirically set to 0.2.
- D. Implementation: Each final-layer forest produces a probability distribution, which is ensembled across forests for each subject.The ensemble probability is used to diagnose COVID-19 versus CAP.
- D. Implementation: The final subject label is selected using the maximum a posteriori criterion over the ensemble probabilities.
- D. Implementation: The collected dataset contains 1495 COVID-19 patients and 1027 CAP patients for the classification experiments.
A. Competing Methods
The study compares AFS-DF with four widely adopted classifiers using normalized location-specific CT features and five-fold cross-validation. Performance is assessed with ACC, SEN, SPE, and AUC, based on confusion-matrix quantities.
- A. Competing Methods: The competing methods are logistic regression, SVM with a radial basis function kernel, random forest, and neural network.These classifiers are compared with AFS-DF on extracted location-specific features.
- A. Competing Methods: All extracted location-specific features are normalized to center 0 and deviation 1 before evaluation.
- A. Competing Methods: Five-fold cross-validation uses each subset once for testing and the remaining four subsets for training, with results averaged across folds.
- A. Competing Methods: Performance is measured using classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve.
- A. Competing Methods: ACC is computed from true positives and true negatives over all four confusion-matrix outcomes.TP, TN, FP, and FN denote true positive, true negative, false positive, and false negative, respectively.
C. Classification Performance
AFS-DF achieves the strongest reported COVID-19-versus-CAP classification performance among the evaluated methods. Its results are assessed in five-fold cross-validation, and feature-importance analysis highlights surface-area features.
- C. Classification Performance: The evaluation uses five-fold cross-validation for COVID-19-versus-CAP classification and reports ACC, SEN, SPE, and AUC.
- C. Classification Performance: AFS-DF achieves the best ACC, SEN, SPE, and AUC among the compared methods.
- C. Classification Performance: 91.79% ACC is achieved by AFS-DF, exceeding LR at 89.81%, SVM at 89.97%, RF at 89.41%, and NN at 89.53%.The reported ACC improvements over LR, SVM, RF, and NN are 1.98%, 1.82%, 2.38%, and 1.83%, respectively.
- C. Classification Performance: AFS-DF produces the best classification performance among the baseline methods in the ROC-curve comparison.
- C. Classification Performance: Surface-area features have important influence for COVID-19 classification according to feature-importance statistics from the last layer.
IV. DISCUSSION
The discussion section includes comparisons with state-of-the-art methods and presents the study’s limitations and future research directions. A referenced figure contains ROC curves for the evaluated classifiers.
- IV. DISCUSSION: The discussion presents limitations of the current study and possible future research directions.
- IV. DISCUSSION: Figure 3 reports ROC curves for LR, SVM, RF, NN, and AFS-DF in COVID-19-versus-CAP classification.
- IV. DISCUSSION: The section compares AFS-DF with several state-of-the-art methods for COVID-19 classification.
A. Comparison with State-of-the-art Methods
AFS-DF is evaluated against existing COVID-19 classification approaches and feature-selection variants for COVID-19 versus CAP classification. The reported comparisons indicate competitive or superior performance, with adaptive feature selection outperforming Lasso, ElasticNet, and DF.
- Comparison with State-of-the-art Methods: The study compares AFS-DF with state-of-the-art methods developed for COVID-19 classification using chest CT images.The compared methods include CNN-based, ResNet-based, logistic-regression, and random-forest approaches across different COVID-19 tasks and datasets.
- Comparison with State-of-the-art Methods: AFS-DF achieves competitive classification performance for COVID-19 patient identification in Table III.The authors attribute this result to AFS-DF's use of high-level discriminative representations of extracted features.
- Comparison with State-of-the-art Methods: AFS-DF achieves the best classification performance compared with Lasso, ElasticNet, and DF in Table IV.The comparison evaluates two conventional feature-selection methods and a deep-forest variant without the feature-selection block.
- Comparison with State-of-the-art Methods: AFS-DF outperforms Lasso and ElasticNet through high-level feature representation and outperforms DF through adaptive feature selection.The authors present these results as evidence for the effectiveness of the proposed AFS-DF design.
C. Influence of Features
The paper examines whether selected deep features improve conventional classifiers for COVID-19 versus CAP classification. Across LR, SVM, and RF, the AFS-DF-based variants generally outperform their counterparts, especially on ACC and SEN.
- Influence of Features: AFSDF-LR, AFSDF-SVM, and AFSDF-RF are evaluated against LR, SVM, and RF using selected deep features.The six methods are compared for COVID-19 versus CAP classification.
- Influence of Features: AFSDF-LR, AFSDF-SVM, and AFSDF-RF outperform LR, SVM, and RF in most evaluation metrics.The proposed variants consistently achieve better results in ACC and SEN.
- Influence of Features: The t-SNE visualization indicates that AFS-DF produces more discriminative features than the original location-specific features and DF features.The figure compares dimensionality-reduced representations from the three feature sets.
D. Limitations and Future Work
The current study validates AFS-DF only for COVID-19 versus CAP classification and relies on handcrafted features derived from prior knowledge. Future work targets other COVID-19 tasks and deep-learning-based feature extraction.
- Limitations and Future Work: The method is validated only on the COVID-19 versus CAP classification task.The authors plan to evaluate it on COVID-19 versus normal and severe-versus-non-severe classification tasks.
- Limitations and Future Work: The study extracts handcrafted features using prior knowledge rather than features learned by deep learning.The authors expect deep-learning-based features to support further performance improvement in future work.
- Limitations and Future Work: AFS-DF achieves superior COVID-19 classification performance on the collected dataset compared with several existing methods.The dataset contains 1495 COVID-19 cases and 1027 CAP cases.