Optimization of deep neural network for multiclassification of Pneumonia

Figures & data

Figure 1. Automated pneumonia identification using CNN models.

Figure 2. (A) X-Ray- Normal/Healthy lung (b)X-Ray- Abnormal/pneumonia affected lung.

Figure 3. Fire protection module in SqueezeNet.

Figure 4. Schematic of SqueezeNet.

Figure 5. Residual learning block.

Figure 6. Schematic representation of ResNet-50 architecture.

Figure 7. EfficientNet-b0 schematic.

Figure 8. Multiclassification of chest X-ray for pneumonia identification using CNN.

Figure 9. Chest-xray classification-1^st stage results using (a) EfficientNet-b0, (b) SqueezeNet, (c) ResNet-50.

Figure 10. Confusion matrix for stage-1 chest X-ray classification using EfficientNet-b0 with accuracy 99%.

Figure 11. Comparison of results of 1st stage for all the 3-pretrained deep neural network(SqueezeNet, ResNet-50, EfficientNet-b0).

Figure 12. Confusion matrix for 2^nd stage- chest X-ray classification using EfficientNet-b0.

Figure 13. Comparison of results of 2^nd stage for all the 3-pretrained deep neural network(SqueezeNet, ResNet-50, EfficientNet-b0).

Figure 14. Chest-X ray classification- 2^nd stage results using (a) ResNet-50, (b) SqueezeNet, (c) EfficientNet-b0.

Table 1. Performance metrics of EfficientNet-b0 & ResNet-50.

1^st stage: X-ray classification using EfficientNet-b0 (solver-adam,batch size-8)		2^nd stage: X-ray classification using ResNet-50 (solver-sgdm,batch size-6)
Accuracy	99 %	97 %
Precision	100 %	94.87 %
Recall	97.14 %	97.36 %
f1 score	98.54 %	96.04 %
Specificity	98.48 %	98.36 %
FNR	1.51%	1.6 %

Figure 15. Workflow of pneumonia diagnosis app.

Data availability statement

The data used to support the findings of this study are included within the article.