[Paper Review] COVID-CXNet: Detecting COVID-19 in Frontal Chest X-ray Images using Deep Learning

Arman Haghanifar, Mahdiyar Molahasani Majdabadi+3 more|arXiv (Cornell University)|Jun 16, 2020

COVID-19 diagnosis using AIMedicine32 citations

TL;DR

COVID-CXNet fine-tunes CheXNet on a COVID-19 CXR dataset with lung-segmentation and image enhancements, achieving high binary-class accuracy and precise pneumonia localization via Grad-CAM heatmaps.

ABSTRACT

One of the primary clinical observations for screening the infectious by the novel coronavirus is capturing a chest x-ray image. In most of the patients, a chest x-ray contains abnormalities, such as consolidation, which are the results of COVID-19 viral pneumonia. In this study, research is conducted on efficiently detecting imaging features of this type of pneumonia using deep convolutional neural networks in a large dataset. It is demonstrated that simple models, alongside the majority of pretrained networks in the literature, focus on irrelevant features for decision-making. In this paper, numerous chest x-ray images from various sources are collected, and the largest publicly accessible dataset is prepared. Finally, using the transfer learning paradigm, the well-known CheXNet model is utilized for developing COVID-CXNet. This powerful model is capable of detecting the novel coronavirus pneumonia based on relevant and meaningful features with precise localization. COVID-CXNet is a step towards a fully automated and robust COVID-19 detection system.

Motivation & Objective

• Collect and curate the largest public COVID-19 CXR dataset from multiple sources to enable robust detection.
• Evaluate multiple CNN architectures including pretrained ImageNet models and CheXNet-based backbones for COVID-19 detection on CXR data.
• Improve localization of pneumonia findings using lung segmentation and image enhancement to guide model focus.
• Assess model interpretability with Grad-CAM and LIME visualizations to verify decision regions.
• Explore multiclass and hierarchical classifications to differentiate COVID-19 pneumonia from non-COVID pneumonia and normal cases.

Proposed method

• Assemble a large public dataset of frontal CXR images (COVID-19 and normal) from multiple sources."
• Preprocess with normalization, resizing to 320x320, and extensive augmentation (including zoom and brightness).
• Apply a U-Net based segmentation to extract lung ROIs and use ROI-cropped inputs for models.
• Fine-tune CheXNet-derived DenseNet backbones (DenseNet-121) on COVID-19 CXRs with dropout and label smoothing.
• Incorporate image enhancement (CLAHE, BEASF) and lung-segmentation to improve localization of abnormalities.
• Experiment with base CNNs, ImageNet pretrained networks (DenseNet, ResNet), and CheXNet-based backbones; evaluate with accuracy, F1 for positive class, and Grad-CAM/LIME visualizations.

Experimental results

Research questions

• RQ1 Can a CheXNet-based backbone fine-tuned on COVID-19 CXR data accurately detect COVID-19 pneumonia in frontal CXRs?
• RQ2 Do lung-segmented ROI inputs and image enhancements improve localization and classification performance compared to non-segmented inputs?
• RQ3 How do pretrained ImageNet models compare to CheXNet-based models for COVID-19 CXR detection in terms of accuracy and interpretability?
• RQ4 What is the effectiveness of hierarchical and multiclass setups (Normal, CAP, CP) for discrimination among normal, non-COVID pneumonia, and COVID-19 pneumonia?

Key findings

• COVID-CXNet achieves 99.04% accuracy and 0.96 F1 on binary COVID-19 vs Normal classification.
• Incorporating lung segmentation (ROI) improves localization of pneumonia findings via Grad-CAM heatmaps compared to non-segmented inputs.
• Compared to base models and ImageNet pretrained networks, CheXNet-based COVID-CXNet provides better PM localization and competitive accuracy.
• Multiclass (Normal, CAP, CP) classification yields 81.04% accuracy, with F-scores 0.85 (CAP) and 0.76 (CP).
• Hierarchical multiclass approach improves overall accuracy to 87.21%, with F-scores 0.92 (CP) and 0.85 (CAP).
• Visualization analyses (Grad-CAM) indicate the model focuses on relevant lung regions, though occasional attention to extraneous text regions can occur without preprocessing.