[Paper Review] A Cascaded Learning Strategy for Robust COVID-19 Pneumonia Chest X-Ray Screening

Chun‐Fu Yeh, Hsien-Tzu Cheng+21 more|arXiv (Cornell University)|Apr 24, 2020

COVID-19 diagnosis using AI8 references36 citations

TL;DR

The paper proposes a three-stage cascaded learning framework (lung segmentation, pneumonia vs non-pneumonia, and COVID-19 vs non-COVID-19) to robustly screen COVID-19 pneumonia from chest X-rays, using incremental learning and explainable heatmaps.

ABSTRACT

We introduce a comprehensive screening platform for the COVID-19 (a.k.a., SARS-CoV-2) pneumonia. The proposed AI-based system works on chest x-ray (CXR) images to predict whether a patient is infected with the COVID-19 disease. Although the recent international joint effort on making the availability of all sorts of open data, the public collection of CXR images is still relatively small for reliably training a deep neural network (DNN) to carry out COVID-19 prediction. To better address such inefficiency, we design a cascaded learning strategy to improve both the sensitivity and the specificity of the resulting DNN classification model. Our approach leverages a large CXR image dataset of non-COVID-19 pneumonia to generalize the original well-trained classification model via a cascaded learning scheme. The resulting screening system is shown to achieve good classification performance on the expanded dataset, including those newly added COVID-19 CXR images.

Motivation & Objective

• Motivate robust COVID-19 screening from chest X-ray images despite limited COVID-19 data.
• Leverage a large open pneumonia dataset to generalize a base classifier.
• Develop a three-stage cascaded learning framework to go from normal/pneumonia to COVID-19 vs non-COVID-19.
• Provide interpretable heatmaps to explain model decisions and facilitate clinical adoption.

Proposed method

• Stage 1: Use a U-Net to predict lung masks and filter non-informative regions in CXR images.
• Stage 2: Train a DenseNet-121 pneumonia classifier on masked images to separate Normal vs Pneumonia (COVID-19 and non-COVID-19 combined as positive).
• Stage 3: Use Stage 2 outputs to mask the input for a fine-grained COVID-19 vs non-COVID-19 pneumonia classifier, with Grad-CAM/Guided-Grad-CAM explanations.
• Incremental learning in Stage 2 to incorporate new COVID-19 data while preserving performance on original data via a knowledge-distillation term in the loss (L_D = (1/|D|) sum CE losses + λ sum KL losses over D_o).
• Stage-wise heatmaps (H_i^2 and H_i^3) provide explainable cues by highlighting lung regions and COVID-19-specific patterns, respectively.

Experimental results

Research questions

• RQ1 Can a cascaded training strategy improve sensitivity and specificity for COVID-19 detection on chest X-rays given limited COVID-19 data?
• RQ2 Does incorporating lung segmentation and incremental learning improve robustness across open and clinical datasets?
• RQ3 Do stage-wise heatmaps enable clinically interpretable explanations for COVID-19 predictions?
• RQ4 How does the proposed method compare to existing COVID-19 X-ray screening approaches on open benchmarks and clinical data?

Key findings

• Stage 2 pneumonia classifier with lung masking yields high discrimination between Normal and Pneumonia on open data (AUC around 97, with high sensitivity and specificity).
• Stage 3 COVID-19 classifier achieves strong performance when trained with open data and later fine-tuned on clinical data (AUCs approaching 99.7 in open validation for COVID-19; high sensitivity and specificity in several settings).
• Incorporating clinical data (NTUH, TMUH, NHIA) improves performance and yields clinically relevant results (e.g., Stage 3 AUCs near 99.7 on open+clinical validation).
• The method demonstrates early COVID-19 detection capabilities relative to RT-PCR in a case study, detecting a subset of cases days before confirmation.
• Compared with COVID-Net on open data splits, the proposed method shows higher sensitivity for Normal and COVID-19 predictions in key evaluations.
• The integrated screening platform has been deployed in Taiwan’s NHIA and CDC systems and demonstrates clinically relevant performance on NHIA data.