Effect of Pre-processing of CT Images on the Performance of Deep Neural Networks Based Diagnosis of COVID-19

Revelo Luna, David Armando; Mejía Manzano, Julio Eduardo; Munoz Chaves, Javier Andres


COVID-19 disease is considered a new challenge around the world. Molecular testing is frequently used, aiming an early detection. However, due to its complexity in the sampling protocol and delay diagnostic, it makes critical the time to decisions on treatment or clinical interventions. In this work, the deep learning technique was adopted to evaluate the performance of 4 systems based on convolutional neural networks (VGG16, VGG19, ResNet50, and MobileNet) to support the diagnosis of COVID-19. CNN models were trained and tested using 340 CT images of patients diagnosed with COVID-19, and the same numbers of images of patients without viruses, 1700 images were obtained for each class using data-augmentation. On these images sets two types of pre-processing were performed normalization and entropy. The parameters: accuracy, recall, precision, and F1Score were used as evaluation metrics. The study found that the best performance in the classification of CT images of patients with COVID-19 was obtained by the MobileNet network with normalization pre-processing attaining 98.04% accuracy. These findings suggest that the type of pre-processing influences CNN's performance strongly. So as a guideline for future development, attention must be paid to implementing pre-processing modules dedicated to highlighting the features of CT images image of COVID-19 positives cases to improve the CNN performance.


Convolutional neural networks, Deep learning, Entropy, Normalization, Transfer learning

Full Text: PDF (downloaded 217 times)


  • There are currently no refbacks.
This abstract viewed 659 times