PaperlayerView exact revision on arXiv

Source-linked AI summary

CoroNet: A deep neural network for detection and diagnosis of COVID-19 from chest x-ray images

Asif Iqbal Khan, Junaid Latief Shah, Mudasir Bhat

arXiv:2004.04931v3eess.IVcs.LGstat.ML

TL;DR

Rapidly increasing COVID-19 cases motivated CoroNet, a deep-learning model for detecting infection from chest X-rays. It achieved reported accuracies of 89.5%, 94.59%, and 99% on 4-class, 3-class, and binary classification tasks, respectively.

  • Problem

    The study addresses the need to detect COVID-19 from chest X-rays amid rapidly increasing case growth.

  • Method

    CoroNet uses a deep-learning convolutional neural network to detect COVID-19 infection from chest X-ray images.

  • Results

    89.5%, 94.59%, and 99% accuracy were reported for 4-class, 3-class, and binary classification tasks, respectively.

  • Takeaways & Limitations

    CoroNet can still be beneficial for radiologists and health experts.

  • Takeaways & Limitations

    CoroNet still needs clinical study and testing.

Abstract

from arXiv · show

Background and Objective The novel Coronavirus also called COVID-19 originated in Wuhan, China in December 2019 and has now spread across the world. It has so far infected around 1.8 million people and claimed approximately 114,698 lives overall. As the number of cases are rapidly increasing, most of the countries are facing shortage of testing kits and resources. The limited quantity of testing kits and increasing number of daily cases encouraged us to come up with a Deep Learning model that can aid radiologists and clinicians in detecting COVID-19 cases using chest X-rays. Methods In this study, we propose CoroNet, a Deep Convolutional Neural Network model to automatically detect COVID-19 infection from chest X-ray images. The proposed model is based on Xception architecture pre-trained on ImageNet dataset and trained end-to-end on a dataset prepared by collecting COVID-19 and other chest pneumonia X-ray images from two different publically available databases. Results and Conclusion CoroNet has been trained and tested on the prepared dataset and the experimental results show that our proposed model achieved an overall accuracy of 89.6%, and more importantly the precision and recall rate for COVID-19 cases are 93% and 98.2% for 4-class cases (COVID vs Pneumonia bacterial vs pneumonia viral vs normal). For 3-class classification (COVID vs Pneumonia vs normal), the proposed model produced a classification accuracy of 95%. The preliminary results of this study look promising which can be further improved as more training data becomes available. Overall, the proposed model substantially advances the current radiology based methodology and during COVID-19 pandemic, it can be very helpful tool for clinical practitioners and radiologists to aid them in diagnosis, quantification and follow-up of COVID-19 cases.

INTRODUCTION

The introduction motivates chest-radiography-based COVID-19 screening amid shortages of testing resources and presents CoroNet to distinguish COVID-19 pneumonia from bacterial and viral pneumonia. It frames the model as an aid for triage, quantification, and follow-up rather than a complete replacement for existing testing.

  • Motivation: Rapidly increasing COVID-19 cases have pushed health systems toward collapse while creating shortages of ventilators and testing kits.
  • Motivation: Effective screening is critical for isolating and treating positive patients, but existing chest-radiography approaches require rigorous testing before production use.
  • Motivation: Chest radiography may provide a fast, widely available approach that can analyze multiple cases and support detection, quantification, and follow-up.
  • Contribution: CoroNet is proposed as a deep convolutional neural network for detecting COVID-19 from chest X-rays and classifying bacterial, viral, and COVID-19 pneumonia.
  • Contribution: The authors also implement binary and 3-class versions, compare them with prior studies, and position the model as support for doctors’ triage and expert review.

DATASET

Because no appropriately sized COVID-19 dataset was available, the authors assembled a chest X-ray dataset from two public databases. They resized the images and applied random under-sampling to address class imbalance.

  • Dataset construction: The dataset was created by collecting chest X-ray images from two publicly available databases because no appropriately sized dataset was available for the study.COVID-19 images came from an open-source repository, while bacterial pneumonia, viral pneumonia, and normal images came from Kaggle.
  • Dataset composition: Around 290 COVID-19 chest radiography images were available in the open database when the paper was written.The repository was compiled from authentic sources and updated regularly.

METHODOLOGY

The methodology section outlines the proposed technique’s work methodology, model architecture, implementation, and training, with the methodology illustrated in Figure 2.

  • The section presents the work methodology for the proposed technique.
  • It discusses the model architecture, implementation, and training.
  • The work methodology is also illustrated in Figure 2.

Convolutional Neural Network (CNN)

CNNs use stacked layers to learn increasingly complex image features through local connections and weight sharing. In CoroNet, convolution, activation, pooling, and fully connected layers support feature extraction, dimensionality reduction, and Softmax-based class prediction.

  • Convolution Layer: Convolutional layers apply learnable filters to local image regions, detecting shared features and mapping them into feature maps.For input matrix I and filter K, convolution produces feature map F, denoted I*K.
  • Activation Function: ReLU introduces non-linearity by outputting 0 for negative inputs and the raw input otherwise.The convolutional-layer output is passed to an activation function before subsequent processing.
  • Subsampling (Pooling) Layer: Pooling layers downsample feature maps, reducing spatial size and network parameters while summarizing local neuron regions.Max pooling selects the maximum value in each input region; average pooling and L2-norm pooling are alternatives.
  • Fully Connected Layer: Fully connected layers connect every neuron to the next layer and produce class scores used for prediction.CoroNet uses Softmax, which computes a probability distribution over output classes whose values sum to 1.
  • Optional Layers: CNN architectures may additionally use batch normalization to improve training time and dropout to address overfitting.These layers are optional additions to the main CNN layers.

Model Architecture and Development

CoroNet is a 71-layer CNN based on Xception and tailored to detect COVID-19 from chest X-ray images. Its design combines ImageNet pre-training, depthwise separable convolutions, residual connections, dropout, and fully connected layers.

  • CoroNet is a CNN architecture tailored for detecting COVID-19 infection from chest X-ray images and based on Xception.
  • 71 layers deep, Xception uses depthwise separable convolution layers with residual connections instead of classical convolutions.Depthwise separable convolution uses point-wise and channel-wise spatial operations, reducing operations by a factor proportional to 1/k.
  • Residual connections allow gradients to flow directly through the network, helping avoid vanishing gradients.The output of a weight-layer series is added to the original input before nonlinear activation.
  • CoroNet uses Xception as its base model with a dropout layer and two fully connected layers added at the end.
  • 33,969,964 parameters in total, with 33,969,964 trainable and 54528 non-trainable parameters, were reported for CoroNet.The architecture details, layer-wise parameters, and output shapes are presented in Table II.
  • Transfer Learning was used to initialize model parameters because the training data was limited and overfitting was a concern.

Implementation and Training

CoroNet was implemented in three classification scenarios based on an ImageNet-pretrained model and retrained end-to-end on the prepared dataset. The main 4-class model was evaluated using four-fold cross-validation.

  • Model scenarios: The study implemented 4-class, 3-class, and binary 2-class CoroNet scenarios for detecting COVID-19 from chest X-ray images.The 4-class model distinguishes COVID-19, Normal, Pneumonia-bacterial, and Pneumonia-viral; the other models modify this main multi-class model.
  • Implementation and training: CoroNet was implemented in Keras on TensorFlow 2.0, initialized with ImageNet pre-training, and retrained end-to-end on the prepared dataset.Training used Adam with learning rate 0.0001, batch size 10, and 80 epochs, with data shuffled before each epoch.
  • Cross-validation: 4-fold cross-validation assessed the main 4-class model by repeatedly training on three sets and validating on the remaining set.The training set was randomly divided into 4 equal sets, and the final performance was averaged across folds.
  • Experimental setup: Experiments and training ran on a Google Colaboratory Ubuntu server equipped with a Tesla K80 graphics card.Accuracy and loss plots for training and validation over epochs were shown for Fold 4.

RESULTS

CoroNet achieved 89.6% average accuracy across the four-class task, with strong COVID-19 class performance. Combining bacterial and viral pneumonia into one class increased accuracy to 94.59%.

  • Four-class classification: 89.6% average accuracy was achieved by CoroNet across the four-class classification task.Accuracy, precision, recall, and F-measure were computed across folds and summarized as average results.
  • Four-class classification: 96.6% average accuracy, 93.17% precision, 98.25% recall, and 95.6% F-measure were reported for the COVID-19 class.The class-wise performance of CoroNet was presented separately from overall four-class performance.
  • Four-class classification: Non-COVID-19 pneumonia classes performed comparatively lower than the other two classes and contributed to lower overall accuracy.The bacterial and viral pneumonia categories were identified as the weaker-performing classes.
  • Three-class classification: 94.59% overall accuracy was obtained after combining bacterial and viral pneumonia into one class, up from 89.5%.The researchers fine-tuned the existing four-class model for three classes and tested it on 29 COVID-19, 72 normal, and 120 pneumonia cases.

Classification accuracy on Dataset-2

CoroNet was evaluated on a second dataset containing normal, pneumonia, and COVID-19 chest X-ray images. After slight modification and fine-tuning, the model achieved 90% overall accuracy on Dataset-2.

  • Dataset-2 contains around 500 normal, 500 pneumonia, and 157 COVID-19 chest X-ray images.
  • The dataset uses the same COVID-19 X-ray images as the prepared dataset, while normal and pneumonia images came from Wang et al.'s ChestX-ray database.
  • 90% overall accuracy was achieved after slight modification and fine-tuning of CoroNet on Dataset-2.The results are illustrated in Table VI, with a corresponding confusion matrix in Figure 7.

DISCUSSION

The discussion reports that CoroNet, an Xception-based model, performed strongly across COVID-19 chest X-ray classification tasks. The authors emphasize reduced missed cases while noting that broader patient data are needed for further improvement.

  • Model performance: The Xception-based model was tested on two datasets and performed exceptionally well on both.The study proposed a deep model based on Xception architecture for detecting COVID-19 from chest X-ray images.
  • Model performance: CoroNet achieved accuracies of 89.5%, 94.59% and 99% for 4-class, 3-class and binary classification, respectively.The model also achieved 90% accuracy on dataset-2.
  • Clinical relevance: Higher recall for COVID-19 cases indicates fewer false negatives, aligning with the study’s aim of minimizing missed cases.The discussion identifies low false-negative counts as an encouraging result.
  • Comparison with prior work: The authors report that CoroNet’s results were superior to other automated COVID-19 chest X-ray studies.The comparison includes COVID-Net’s 83.5% accuracy for four classes and other deep-learning models’ reported results.
  • Limitations: Further improvement requires more patient data, including symptomatic and asymptomatic cases, for deeper analysis.The discussion presents limited available patient data as a study limitation.

CONCLUSION

CoroNet is a convolutional neural network for identifying COVID-19 cases from chest X-ray images, trained and tested on a small dataset assembled from public databases. It achieved promising results but requires more data and clinical testing before broader use.

  • CoroNet is a convolutional neural network designed to detect COVID-19 cases from chest radiography and chest X-ray images.
  • The model was trained and tested on a small dataset of a few hundred images containing COVID-19 and various pneumonia cases from public databases.
  • CoroNet is computationally less expensive and achieved promising results on the prepared dataset.
  • Performance could improve with more training data, but CoroNet still needs clinical study and testing.
  • With higher accuracy and sensitivity for COVID-19 cases, CoroNet could benefit radiologists and health experts in understanding critical aspects of COVID-19.
Loading 2004.04931v3…