Deep Learning Based Pulmonary Edema Detection: Performance Evaluation of Optimizers

Abstract

Pulmonary edema is a life-threatening condition that is caused by fluid accumulation in the lungs, remains a diagnostic challenge, particularly in resource limited healthcare settings. This research presents a deep learning-based approach for automated detection of pulmonary edema using frontal chest X-rays from the NIH Chest X-ray14 dataset. The study proposes a methodology for pulmonary edema detection using the custom 7-layer convolutional neural network (CNN) in combination with different deep learning and hybrid algorithms. A dataset of total 4,606 images was used to train, validate and evaluate the algorithms. A total of 30 different deep learning and hybrid algorithms were experimented on different thresholds of 0.3, 0.4 and 0.5. Total 6 algorithms on threshold 0.5 were selected for further experimentation using the proposed methodology namely DenseNet121, ANN, DenseNet121+VGG16, DenseNet121+VGG19 and DenseNet121+NasNet. To enhance model performance and generalization, various preprocessing techniques, hyperparameters and optimizers including SGD, RMSProp, Adam, Nadam, AdaDelta and AdaGrad were applied on threshold 0.5. The DenseNet121+NasNet model demonstrated an excellent performance at 0.5 threshold, achieving 94.55% accuracy, 99.27% precision, 89.77% recall, an F1-score of 0.9428, and an AUC of 0.9948. The results of the proposed methodology and other existing methodologies are also compared using AUC. The comparison clearly shows that our proposed methodology results are excellent with highest AUC of 0.9948. The findings clearly show that our proposed methodology is an accurate and reliable solution to develop a scalable diagnostic tool for pulmonary edema, supporting medical professionals in underprivileged regions with a shortage of skilled radiologists.