[Paper Review] Adaptive Feature Selection Guided Deep Forest for COVID-19 Classification with Chest CT

Liang Sun, Zhanhao Mo+15 more|arXiv (Cornell University)|May 7, 2020

COVID-19 diagnosis using AI27 references18 citations

TL;DR

This paper proposes an Adaptive Feature Selection guided Deep Forest (AFS-DF) model for COVID-19 classification from chest CT scans, leveraging location-specific features and a deep forest architecture to learn high-level representations while adaptively reducing feature redundancy. The method achieves 91.79% accuracy, 93.05% sensitivity, 89.95% specificity, and 96.35% AUC on a dataset of 1,495 COVID-19 and 1,027 CAP cases, outperforming baseline methods including Lasso, ElasticNet, and standard Deep Forest.

ABSTRACT

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.

Motivation & Objective

• To address the challenge of limited medical imaging data in training robust models for COVID-19 classification.
• To improve classification performance by leveraging high-level representations from location-specific features in chest CT scans.
• To reduce feature redundancy and enhance model interpretability through a task-driven, adaptive feature selection mechanism integrated with the deep forest framework.
• To develop a machine learning model that outperforms traditional methods like Lasso, ElasticNet, and standard Deep Forest on small-scale medical imaging datasets.

Proposed method

• Extract location-specific features from chest CT images using handcrafted methods informed by anatomical prior knowledge.
• Employ a deep forest model to learn hierarchical, high-level representations from the extracted features, enabling effective representation learning with limited training data.
• Introduce an adaptive feature selection strategy that computes feature importance from trained decision trees in each deep forest layer and prunes less relevant features iteratively.
• Integrate feature selection and classification into a unified framework, where feature pruning is guided by the model's internal performance feedback during training.
• Train the AFS-DF model end-to-end on a dataset of 1,495 COVID-19 and 1,027 CAP cases to optimize for classification performance.
• Validate the method using 10-fold cross-validation and compare performance against Lasso, ElasticNet, and standard Deep Forest using standard metrics.

Experimental results

Research questions

• RQ1 Can an adaptive feature selection mechanism improve the performance of a deep forest model in classifying COVID-19 from community-acquired pneumonia using chest CT scans?
• RQ2 Does the integration of high-level representation learning via deep forest architecture enhance classification accuracy and robustness on small medical imaging datasets?
• RQ3 How does the proposed AFS-DF model compare to traditional machine learning methods such as Lasso and ElasticNet in terms of sensitivity, specificity, and AUC for COVID-19 diagnosis?
• RQ4 To what extent do the selected deep features from the AFS-DF model improve the performance of conventional classifiers like LR, SVM, and RF?

Key findings

• The AFS-DF model achieved 91.79% accuracy, 93.05% sensitivity, 89.95% specificity, and 96.35% AUC on the test set, outperforming Lasso, ElasticNet, and standard Deep Forest.
• The AFS-DF method improved accuracy by 1.38 percentage points over logistic regression (LR), 1.15 points over SVM, and 1.11 points over random forest (RF) when using the same selected deep features.
• The feature selection mechanism effectively reduced redundancy, as evidenced by more discriminative feature representations in t-SNE visualizations of the last layer features.
• The use of deep features selected by AFS-DF significantly enhanced the performance of conventional models, indicating that the selected features contain high-level, discriminative information.
• The adaptive feature selection process, based on internal forest importance scores, contributed to improved model generalization and robustness on small datasets.