Explainable optimization of deep learning model for COVID-19 detection using chest images

• Published in: Informatics in medicine unlocked 2024, Vol.49, p.101559, Article 101559
• Main Authors: Riyadi, Slamet, Pramudya, Eka Nova, Damarjati, Cahya, Molina Lopez, Jose Manuel, Herrero, Jesus Garcia
• Format: Article
• Language: English
• Subjects: Activation maps; COVID-19; Deep learning; Grad-CAM; Optimizer; VGG; X-ray
• ISSN: 2352-9148; 2352-9148
• DOI: 10.1016/j.imu.2024.101559

The COVID-19 pandemic has now become endemic, yet it remains essential to identify and diagnose the virus. The X-ray images on chest evaluations have utilized numerous deep-learning techniques and optimization algorithms. The Visual Geometry Group (VGG) is a widely recognized deep-learning architecture for COVID-19 detection. This architecture incorporates various optimization algorithms, such as Adadelta, Adam, Adamax, Stochastic Gradient Descent (SGD), Adagrad, Adam, and Root Mean Square Propagation (RMSprop), to ascertain the most favorable learning parameters. Currently, researchers evaluate the performance of the deep learning and optimization algorithm to detect COVID-19 by computing the statistical parameters, i.e., the accuracy, precision, recall, specificity, and F-1 score. However, the way in which deep learning and optimization algorithms work to predict the outcome still needs a comprehensive explanation. This study aims to explain how deep learning and optimization predict COVID-19 or normal chest using visual and quantitative analysis of a Gradient Class Activation Mapping (Grad-CAM) and statistical evaluation. The methodology involved data collection of chest images from a public data set, preprocessing, training-testing, and result analysis on statistical parameters and Grad-CAM heatmap. The evaluation of statistical parameters yielded favorable outcomes from all the optimization techniques, with an average accuracy of up to 99 %. It was then followed by visually observing the Grad-CAM activation heatmap, which indicated crucial regions in the images that influenced the model's prediction outcomes. Grad-CAM allows the visualization and quantification of the activation map in each optimization. It reveals distinct activation maps for the lung area in both COVID-19 and normal pictures. The visual observation was also confirmed by quantitative analysis of Grad-CAM using root mean square error (RMSE) of correlation between each optimization algorithm. The researcher compares two heatmaps from different optimization methods by computing the RMSE of each heatmap pixel value. The RMSE confirmed that the heatmap supports the explanation of the optimization performance. In conclusion, this research explained how deep learning and optimization algorithms predict COVID-19 from the chest images.