Can AI help in screening Viral and COVID-19 pneumonia?

I. INTRODUCTION

• RT-PCR is manual, laborious, time-consuming, has a reported positivity rate of 63%, and costs about USD 120–130 per kit, with results delayed by supply shortages.The passage also notes that PCR machines require specialized biosafety labs and that delays can allow infected patients to interact with healthy patients.Source: §I · p. 2
• Although chest X-rays may support early screening, overlapping appearances across viral pneumonias and other lung diseases make COVID-19 difficult to distinguish and can lead to misdiagnosis.A non-COVID viral pneumonia may be falsely labeled as suspected COVID-19, delaying treatment and increasing exposure risk.Source: §I · p. 2
• Deep CNNs can learn image features, and transfer learning allows pretrained networks to be retrained using comparatively few images, motivating their use for chest-X-ray analysis.Prior applications cited include pneumonia detection and classification, thoracic-disease identification, and image segmentation.Source: §I · p. 2
• Recent deep-learning COVID-19 X-ray studies often used small datasets, limiting confidence in generalization; rigorous transfer-learning evaluation on large COVID and non-COVID datasets was scarce [40] [44] [45] [46] [47] [49] [51] [53] [54] [55] [56].The introduction contrasts this gap with prior approaches including conventional CNNs, capsule networks, and modified or fine-tuned pretrained models.Source: §I · pp. 2–3
• The authors created a public dataset of 1,579 normal, 1,485 viral-pneumonia, and 423 COVID-19 chest X-rays and evaluated eight pretrained networks across two classification schemes, with and without augmentation.One scheme distinguishes normal from COVID-19 images; the other distinguishes normal, viral pneumonia, and COVID-19 pneumonia.Source: §I · p. 3

II. METHODOLOGY

• The study used transfer learning to adapt pretrained CNNs to a smaller dataset, replacing the Softmax and classification layers and reusing pretrained weights.The passage describes this as a way to reduce training time and avoid requiring a large dataset.Source: §II · p. 3
• The authors created a Kaggle database and made the trained models available to researchers.The passage says the database was intended to make the dataset publicly available.Source: §II · p. 3

A. DATABASE DESCRIPTION

• The database used AP/PA chest X-rays and combined six sub-databases; the authors assembled the COVID-19 images, while normal and viral pneumonia images came from public Kaggle databases.The COVID-19 collection included material from public databases and published articles.Source: A. DATABASE DESCRIPTION · p. 3
• The assembled dataset comprised 423 COVID-19, 1,485 viral pneumonia, and 1,579 normal chest X-ray images.The COVID-19 images were assembled from multiple sources; normal and viral pneumonia images came from publicly available databases.Source: A. DATABASE DESCRIPTION · pp. 3–4
• The SIRM source listed 384 COVID-19 radiographic images, including 94 chest X-rays; the study’s COVID-19 sub-database contained 423 AP/PA images overall.The SIRM collection also included 290 CT images; its reported case count was 71 confirmed cases as of 10 May 2020.Source: A. DATABASE DESCRIPTION · p. 3
• From the Cohen, Morrison, and Dao GitHub database, the authors used 134 COVID-19 chest X-ray images, distinct from images they collected from articles.That database reported 250 COVID-19 chest X-ray images and 25 COVID-19 lung CT images.Source: A. DATABASE DESCRIPTION · pp. 3–4
• The authors added 60 COVID-19 chest X-rays from 43 articles and 32 from Radiopaedia, excluding images already listed in the GitHub database.They described collecting and indexing images from articles and online sources, then checking for duplication.Source: A. DATABASE DESCRIPTION · p. 4
• Normal and viral pneumonia images came from RSNA and Kaggle sources; a Spanish physician also shared 103 images covering 50 cases.The Kaggle chest X-ray collection contained 5,247 images, including 1,345 viral pneumonia and 1,341 normal images; the study used Kaggle images to create its normal and viral sub-databases.Source: A. DATABASE DESCRIPTION · p. 4

B. CNN Model Selection

• Eight pre-trained CNNs were evaluated: MobileNetv2, SqueezeNet, ResNet18, Inceptionv3, ResNet101, CheXNet, VGG19, and DenseNet201.The study compared three shallow with five deep networks, used two ResNet variants to examine depth in similar architectures, and tested whether X-ray-pretrained CheXNet outperformed other deep networks.Source: B. CNN Model Selection · p. 4
• Models were trained with SGD momentum (learning rate 10^-3, momentum 0.9), mini-batches of 16, and 20 epochs; fivefold cross-validation was averaged for final evaluation outputs.The averaged outputs included ROC curves, confusion matrices, and evaluation metrics.Source: B. CNN Model Selection · p. 4
• Experiments covered two-class and three-class image classification, each using models trained without and with image augmentation.The three-class problem classified normal, viral pneumonia, and COVID-19 images.Source: B. CNN Model Selection · p. 4

B. PREPROCESSING

• Images were resized to 227×227 pixels for SqueezeNet, 224×224 for MobileNetV2, ResNet18/101, VGG19, and DenseNet201, and 299×299 for InceptionV3, then normalized to pretrained-model standards.Source: B. PREPROCESSING · p. 5
• Study 1 used 423 images per class without augmentation, while Study 2 used the full 423 COVID-19, 1,579 normal, and 1,485 viral-pneumonia dataset with training augmentation.Study 1 randomly sampled normal and viral-pneumonia images to match the COVID-19 count; Study 2 augmented training images to create a balanced training set.Source: B. PREPROCESSING · p. 5
• Both studies used stratified 5-fold cross-validation, with 80% training, 20% unseen testing, and 10% of training data reserved for validation.Source: B. PREPROCESSING · p. 5
• Study 2 augmented COVID-19 images six times and normal and viral-pneumonia images once, using 15-degree rotations and 5% horizontal or vertical translations.Source: B. PREPROCESSING · p. 5, Figure 3 · p. 5

C. IMAGE AUGMENTATION

• COVID-19 training images were rotated clockwise and counterclockwise by 5, 10, or 15 degrees and translated horizontally and vertically by −5% to 5%; viral and normal images received translation only.Source: C. IMAGE AUGMENTATION · p. 5
• Study 1 used COVID-19 and normal images, while Study 2 included COVID-19, normal, and viral pneumonia; Table 1 summarizes class counts across training, validation, and testing at each fold.Source: C. IMAGE AUGMENTATION · p. 5

D. INVESTIGATION OF THE DEEP LAYER FEATURES

• Activation maps were normalized to 0–1, and the strongest channels for COVID-19, normal, and viral-pneumonia images were compared with their original images.The strongest channel activated positively on light-left/dark-right edges and negatively on dark-left/light-right edges.Source: D. INVESTIGATION OF THE DEEP LAYER… · p. 6
• CNNs learn edges and colors in early layers, then combine earlier features into more complicated features in deeper layers.Source: D. INVESTIGATION OF THE DEEP LAYER… · p. 6
• Activation mapping highlights image regions contributing to classification and offers a visual explanation of CNN predictions, including successes or failures.The authors propose using activation mapping to illustrate features that distinguish COVID-19 X-rays from the other two classes.Source: D. INVESTIGATION OF THE DEEP LAYER… · p. 6

E. PERFORMANCE EVALUATION MATRIX

• The trained algorithms were validated using 5-fold cross-validation and evaluated with accuracy, sensitivity or recall, specificity, precision (PPV), and F1 score.Source: E. PERFORMANCE EVALUATION MATRIX · p. 6
• Per-class metric values were computed from the overall confusion matrix, which accumulated results across all test folds.Source: E. PERFORMANCE EVALUATION MATRIX · p. 6
• The evaluation covers two-class COVID-19-versus-normal and three-class COVID-19, normal, and viral-pneumonia classification schemes.Source: E. PERFORMANCE EVALUATION MATRIX · p. 6

• The study compared eight pre-trained CNN models, with and without image augmentation, for COVID-19 versus normal classification and for three-class classification including viral pneumonia.Source: §III · p. 6

A. EXPERIMENTAL RESULTS – TWO CLASS PROBLEM

• All evaluated pre-trained models performed very well, with similar weighted-average performance and a small gain from image augmentation.Without augmentation, ResNet18 and CheXNet performed equally; with augmentation, CheXNet and DenseNet201 performed better than the other networks, with marginal differences.Source: A. EXPERIMENTAL RESULTS – TWO CLASS… · p. 6
• CheXNet’s performance in both settings was attributed to its pre-training on a large X-ray image database.The authors note that this pre-training may allow the network to perform well on X-ray image classification without training again on a larger dataset.Source: A. EXPERIMENTAL RESULTS – TWO CLASS… · pp. 6–7
• 99.4% and 99.7% accuracy: CheXNet achieved the highest two-class accuracy without and with image augmentation, respectively.The reported gains over other networks were marginal.Source: A. EXPERIMENTAL RESULTS – TWO CLASS… · p. 6, Table 2 · p. 7
• ROC curves showed comparable performance across all tested networks with and without image augmentation.The authors note that COVID-19 images are significantly different from normal images, and that all tested networks performed well in this classification problem.Source: Figure 5 · p. 7, A. EXPERIMENTAL RESULTS – TWO CLASS… · p. 7

PROBLEM

• Three-class performance was lower than two-class performance on small datasets; augmentation exposed differences between models, with deeper networks typically performing better, though ResNet18 outperformed ResNet101.CheXNet performed well without augmentation, but DenseNet201 surpassed the 121-layer CheXNet when augmentation was used.Source: PROBLEM · pp. 7–8
• DenseNet201 outperformed the other models across performance indices with image augmentation, which significantly improved three-class performance and overall network performance.The ROC comparison also shows DenseNet201 with augmentation increasing overall performance.Source: PROBLEM · p. 8, Figure 6 · p. 8, Table 3 · p. 8
• Only 1 of 423 COVID-19 images was classified as normal in the augmented three-class task; 2 others were classified as viral pneumonia, and no normal images as COVID-19.Four of 1,485 viral-pneumonia images were classified as COVID-19 and 33 as normal; confusion was greater between viral pneumonia and normal than with COVID-19.Source: PROBLEM · pp. 8–9, Figure 7 · p. 9
• DenseNet201’s 14th layer detected features distinguishing normal, COVID-19, and viral-pneumonia images, including a light-focus edge feature described in COVID-19 images.The authors suggest deep-layer features may reveal abnormalities not visible to the human eye.Source: PROBLEM · p. 9, Figure 8 · p. 9
• Three COVID-19 images with little upper-lobe and suprahilar opacity resembled normal X-rays; without an evident light-focus edge, the algorithm could fail and other techniques may be needed.Radiologists saw no or very little COVID-19 signs in two images and very mild infection in the other.Source: PROBLEM · pp. 9–10, Figure 9 · p. 10

IV. CONCLUSION

• Accuracy, precision, sensitivity, and specificity were 99.7%, 99.7%, 99.7%, and 99.55% for normal/COVID-19 classification, and 97.9%, 97.95%, 97.9%, and 98.8% for normal/COVID-19/viral classification.Source: §IV · p. 10
• DenseNet201 outperformed other CNNs with image augmentation, while CheXNet performed best without augmentation; the authors attribute this to its pretraining on a large X-ray database.Source: §IV · p. 10
• The authors report that deep networks performed better than shallow networks, particularly for normal and viral images, while most networks identified COVID-19 with very high sensitivity.Source: §IV · p. 10
• The authors caution that findings from small datasets may not generalize to real applications and state that effective COVID-19 classification requires training on a large dataset.Source: §IV · p. 10
• The authors suggest computer-aided diagnosis could speed early screening and improve screening accuracy amid limited healthcare resources and variability in X-ray images.Source: §IV · p. 10

FUNDING

• Publication funding came from the Qatar National Library, while the work was supported by NPRP12S-0227-190164 from the Qatar National Research Fund.The fund is a member of Qatar Foundation in Doha, Qatar; the authors state that the paper’s statements are solely their responsibility.Source: FUNDING · p. 10