Source-linked AI summary
Deep Learning on Chest X-ray Images to Detect and Evaluate Pneumonia Cases at the Era of COVID-19
Karim Hammoudi, Halim Benhabiles, Mahmoud Melkemi, Fadi Dornaika, Ignacio Arganda-Carreras, Dominique Collard, Arnaud Scherpereel
TL;DR
The paper addresses limited specificity and access to radiographic evidence for COVID-19 screening by developing deep-learning models and health indicators from publicly available chest X-ray datasets. The models detected pneumonia and viral cases, while the indicators estimated infection status and patient risk in simulated scenarios.
Problem
COVID-19 chest imaging findings overlap with other infections, while publicly accessible COVID-19-related imaging datasets and expert radiologists are limited, motivating automated screening support.
Method
The paper trains tailored CNN and RNN models on public chest X-ray datasets to classify pneumonia-related cases and proposes health indicators based on infection rate, age, and comorbidity.
Results
Tailored models exceeded 84% average pneumonia-detection accuracy; DenseNet169 reached 95.72% classification accuracy, while the RNN detected pneumonia in 99.3% of blind COVID-19 test images.
Takeaways & Limitations
Public chest X-ray datasets can support deep-learning screening of viral pneumonia, and the proposed indicators can help evaluate infection and aggravation risk in simulated patient scenarios.
Abstract
from arXiv · showhide
Coronavirus disease 2019 (COVID-19) is an infectious disease with first symptoms similar to the flu. COVID-19 appeared first in China and very quickly spreads to the rest of the world, causing then the 2019-20 coronavirus pandemic. In many cases, this disease causes pneumonia. Since pulmonary infections can be observed through radiography images, this paper investigates deep learning methods for automatically analyzing query chest X-ray images with the hope to bring precision tools to health professionals towards screening the COVID-19 and diagnosing confirmed patients. In this context, training datasets, deep learning architectures and analysis strategies have been experimented from publicly open sets of chest X-ray images. Tailored deep learning models are proposed to detect pneumonia infection cases, notably viral cases. It is assumed that viral pneumonia cases detected during an epidemic COVID-19 context have a high probability to presume COVID-19 infections. Moreover, easy-to-apply health indicators are proposed for estimating infection status and predicting patient status from the detected pneumonia cases. Experimental results show possibilities of training deep learning models over publicly open sets of chest X-ray images towards screening viral pneumonia. Chest X-ray test images of COVID-19 infected patients are successfully diagnosed through detection models retained for their performances. The efficiency of proposed health indicators is highlighted through simulated scenarios of patients presenting infections and health problems by combining real and synthetic health data.
I. INTRODUCTION
The paper frames COVID-19 screening as difficult because testing constraints and nonspecific imaging findings limit practical diagnosis. It proposes deep-learning analysis of publicly available chest X-ray data to detect pneumonia and support clinical evaluation.
- Motivation: COVID-19 screening is arduous because RT-PCR kits may be unavailable and false negatives can contribute to disease propagation.The paper notes that deep CT analysis may predict some true positives when RT-PCR is negative.
- Motivation: Chest imaging findings are nonspecific and overlap with influenza, H1N1, SARS, and MERS.Chest X-rays may also have poor quality because examinations are often performed with patients lying or sitting for hygienic reasons.
- Motivation: Publicly open COVID-19-related CT datasets are limited, while some open chest X-ray datasets are available for deep-learning experiments.The paper presents this availability contrast as a motivation for investigating chest X-ray analysis.
- Contributions: The paper proposes tailored deep-learning models for detecting pneumonia, especially viral cases, and easy-to-apply indicators for infection and patient-status evaluation.The stated aim is to support radiologists and clinicians, not replace medical advice.
- Paper organization: The study experiments with training datasets, image-detection architectures, and analysis strategies, followed by performance evaluation and conclusion.The paper structure assigns experiments to Section 4 and the conclusion to Section 5.
II. RELATED WORK
Related work covers deep-learning approaches for COVID-19 analysis using CT and chest radiography. It positions chest X-rays as widely available resources investigated for upstream screening despite infections being more directly visible in CT images.
- CT-based approaches: Prior studies used deep learning to screen COVID-19 from CT images and to segment and quantify infection in COVID-19 patient scans.The cited studies include a three-group CT classifier and a human-in-the-loop segmentation strategy.
- Chest radiography: Pulmonary infections can be more directly visible in CT than chest X-rays, but chest X-rays remain relevant because they are widespread and often analyzed upstream of CT scans.
- Chest radiography: Prior work also proposed COVID-Net for detecting COVID-19 cases from chest radiography images using open datasets.The derived COVIDx dataset combined 5,941 posteroanterior radiographs from 2,839 patient cases.
- Screening workflow: The paper's workflow links automatic infection detection in chest X-rays with COVID-19 screening, treating detected viral pneumonia as presumptive evidence during an epidemic.
A. CNN-based detection and evaluation of infected patients
The paper treats COVID-19 as a viral disease and uses three-way chest X-ray classification to identify normal, bacterial, or viral pneumonia cases for epidemic-context screening.
- Detection assumption: The classifier predicts whether a query chest X-ray is normal, bacterial, or viral.A COVID-19 infection tested during an epidemic is assumed to have a high probability of being a true positive when the output is Virus.
- Detection assumption: The approach is intended to support early-stage COVID-19 pulmonary-symptom screening, while acknowledging that severe viral infection can lead to secondary bacterial pneumonia.
- Detection assumption: The viral classification assumption is based on COVID-19 being a viral rather than bacterial disease.
1) Tailored CNN models:
The study builds CNN-based classifiers and a dual-use Inception-ResNetV2–RNN model that combines global feature extraction with local image-block predictions.
- Tailored CNN models: Tailored CNN models classify normal, viral-pneumonia, and bacterial cases by outputting a predicted probability for each category.The models use ResNet34, ResNet50, or DenseNet169 backbones with a fully connected classifier head.
- Tailored CNN models: A dual-use model combines a fine-tuned Inception-ResNetV2 feature extractor with a bidirectional LSTM for final image-block categorization.
- Local image analysis: The dual-use model divides each chest X-ray into grid cells and analyzes the resulting sequence of image blocks for local predictions.This second-level analysis is intended to retain image details that resizing may lose.
- Local image analysis: Local predictions form a contamination matrix that supports estimation of health indicators such as a CNN-based infection ratio.The grid discretization is tuned according to the predictive performance of the architecture.
- Test data: Figure 3 provides chest X-ray samples from the test datasets.
2) Data preparation and model inputs:
The study reorganizes a public pediatric chest X-ray dataset into normal, bacterial pneumonia, and viral pneumonia classes, then prepares standardized inputs and a COVID-19 blind test set for model evaluation.
- The source dataset contains 5,863 children’s X-ray images labeled as Normal or Pneumonia, with pneumonia further labeled bacterial or viral.
- The dataset is reorganized into three classes, with 1,345 training images and 148 test images per class.
- Images are resized to a common dimension to fit tailored architectures and keep CNN training computationally manageable.
- The final tailored RNN model first splits each original image before resizing.
- A blind test set contains 145 chest X-ray images from COVID-19-infected patients, but its limited size prevents reliable discrimination from non-COVID-19 viral pneumonia.
B. Estimation of CNN-based health indicators
The paper proposes health indicators that combine CNN-derived infection information with age and comorbidity risks to estimate patient status, while allowing temporal disease progression to be incorporated.
- The measure F adds age, infection-rate, and comorbidity penalty scores, then divides their sum by a critical threshold T.Beyond T, the paper states that recovery is not possible.
- The scoring system must be adjusted by health professionals to match reality.
- The infection-related score S2 can use CNN class probability or a refined sub-image infection rate.The refined formulation divides an image into n = 9 sub-images and assigns S2 = (100/n) × N, where N is the number of virus-infected sub-images.
- Each serious comorbidity receives a penalty of 100, while each moderated disease receives a penalty of 10 in the example.Examples of serious diseases include complicated hypertension, coronary artery disease, dialysis-dependent renal failure, and cancer under treatment.
- The example sets the critical threshold T to 200, based on an assumed patient over 80 with a serious additional disease.The paper states that F ≥1 means the hope of escaping fatality is null, while values from 0 to 1 indicate progressively less hope of recovery.
- The measure can incorporate disease progression by comparing infection rates at two time points and applying aggravation, stability, or remission adjustments.The formulation uses t1 and t2, a stability threshold δ = 20, and bonus or malus terms tied to f(t2).
A. Performance of tailored CNN models
Tailored CNN and RNN models showed strong pneumonia detection, including on heterogeneous COVID-19 X-ray images, while image timing and age distributions support early and adult-case applicability.
- CNN classification: 95.72% average classification accuracy was achieved by DenseNet169 on the Chest X-Ray Images (Pneumonia) dataset.Class accuracies were 97.97% for bacterial, 96.62% for virus, and 92.57% for normal cases.
- COVID-19 test performance: 99.3% of pneumonia cases were detected by the RNN architecture on the blind COVID-19 test set using default settings.The same architecture detected viral infection in 60.64% of cases with majority voting in sequences.
- COVID-19 test performance: The models produced robust pneumonia detection and satisfying viral detection despite radiography images originating from diverse sources.The extracted COVID-19 images came from at least 24 hospitals worldwide.
- Clinical applicability: Because many images were acquired during the first week of symptoms or hospitalization, the models may support early-stage detection when pneumonia is present.The passage states that quasi-totality of pneumonia cases were detected.
- Clinical applicability: Models trained mainly on 5,232 pediatric chest X-ray images may remain relevant for adult pneumonia detection across ages 20 to 80.The conclusion is based on the observed age distribution and near-complete pneumonia detection.
B. Projection with the CNN-based health indicators
The paper reports infection-rate estimates from successive X-ray images and evaluates an additional indicator using synthetic patient data.
- Health indicators: RNN-derived infection rates S2 were estimated from real pairs of successive X-ray images for five COVID-19-infected patients.The paper also gives examples of F calculated for nine patients from synthetic data.
V. CONCLUSION
The study found promising performance from tailored deep learning models for pneumonia detection and explored health indicators for infection and aggravation assessment. It also identifies further validation and data needs before broader COVID-19 discrimination.
- Performance: All tailored models exceeded 84% average accuracy for pneumonia detection on the reorganized Pneumonia dataset.Inception-ResNetV2 produced the minimum false-negative rate, 0.7%, on the blind test set.
- Generalization: Transfer learning from pediatric chest X-ray training to adult infection screening was shown to be efficient in the experiments.
- Health indicators: The study proposed easy-to-apply indicators for evaluating infection rate and aggravation risk in COVID-19 pneumonia scenarios.
- Scope and future work: Future work may distinguish COVID-19 viral pneumonia from non-COVID-19 viral pneumonia when sufficient COVID-19 chest X-ray images become available.The authors state that model reliability must also be cross-checked with RT-PCR and clinical tests before deployment.