A modified deep convolutional neural network for detecting COVID-19 and pneumonia from chest X-ray images based on the concatenation of Xception and ResNet50V2

• Quick detection matters because the introduction says limited diagnostic tools and production constraints slow detection, while rapid detection may reduce other diseases and COVID-19 casualties.Source: §1 · p. 1
• The introduction presents chest X-rays and CT scans as accessible imaging sources for COVID-19 detection, including cases without common symptoms such as fever.Source: §1 · p. 2
• Specialist-led diagnosis uses symptoms and chest radiology images to distinguish COVID-19 from healthy or other-disease cases, but it is costly and time-consuming.Source: §1 · p. 2
• The paper proposes combining two deep convolutional networks, optimizing training parameters, and training for imbalanced datasets to improve COVID-19 detection and reduce false detections.Source: §1 · p. 2
• Prior work cited includes CT-based detection and segmentation, COVID-Net, and transfer-learning classification of chest X-rays into normal and COVID-19 classes, [9], [33], [20],.Source: §1 · p. 2

2.1 Neural Networks

• The concatenated network joins Xception and ResNet50V2 features, then applies a 1*1 convolutional layer before the classifier.The authors describe the added layer as extracting semantic features across channels.Source: §2.1 · p. 3, Figure 1 · p. 3

2.2 Dataset

• The COVID-19 class had 180 images, so the authors allocated data into eight training phases with nearly equal class counts and varying pneumonia cases.Each phase contained 633 cases; the phases shared COVID-19 images and some pneumonia cases while using different additional pneumonia cases.Source: §2.2 · p. 4

2.3 Training phase

• The authors trained ResNet50V2, Xception, and their concatenation using five folds; the concatenated network showed higher accuracy than the other networks.Training used categorical cross-entropy, Nadam with a 1e-4 learning rate, 100 epochs per phase, and data augmentation.Source: §2.3 · pp. 5, 7, Table 2 · p. 7

3 Results

• The validation set comprised 31 COVID-19, 4,420 pneumonia, and 6,851 normal cases; training used 3,783 images across eight phases.In fold 3, 30 COVID-19 cases were used for validation and 150 other cases for training.Source: §3 · p. 7
• The networks began training with ImageNet-pretrained weights, and validation accuracy was monitored after each epoch to select the best-converged version.Source: §3 · p. 7
• Evaluation reported class-level and overall accuracy alongside four class-specific metrics; Tables 3 and 4 and Figure 4 present results and confusion matrices.Figure 4 shows confusion matrices for folds 1 and 3.Source: §3 · p. 8, Table 3 · p. 8, Figure 4 · p. 9
• True positives, false positives, false negatives, and true negatives respectively count correct class detections, incorrect class assignments, missed class images, and correctly excluded non-class images.Source: §3 · p. 8

4 Discussion

• The concatenated network performed better in COVID-19 detection, avoiding false COVID-19 classifications, and overall accuracy.In fold one, it correctly detected 26 of 31 COVID-19 cases and classified 68 of 11,271 other cases as COVID-19.Source: §4 · p. 10
• Because validation included few COVID-19 cases and many other cases, false positives could exceed true positives and lower COVID-19 precision.Source: §4 · p. 10
• The authors say that testing many images makes results more practical, while more COVID-19 patient data may raise detection accuracy.Source: §4 · p. 10

5 Conclusion

• The proposed network concatenates Xception and ResNet50V2 for classifying chest X-rays as normal, pneumonia, or COVID-19.Source: §5 · p. 10
• Training used eight successive phases of 633 images, with nearly equal class counts per phase and different normal and pneumonia images across phases.The training set contained 3,783 images; the remaining images were allocated for evaluation.Source: §5 · p. 10
• 99.50% average accuracy and 80.53% sensitivity for the COVID-19 class, and 91.4% overall accuracy across five folds, were reported.Source: §5 · p. 11

6 Code Availability

• The trained networks are publicly shared on the authors’ GitHub profile.Source: §6 · p. 11
• The code used in the paper is available on the authors’ GitHub profile.Source: §6 · p. 11
• The authors hope the shared work will be useful in future research.Source: §6 · p. 11