Analysis of radiographic images of COVID19 patients for predicting disease development
| Autor: | Ramirez Moreno, Joel |
|---|---|
| Přispěvatelé: | Díaz Honrubia, Antonio Jesús |
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | Archivo Digital UPM Universidad Politécnica de Madrid |
| Popis: | The year 2020 was marked by the arrival of the SARS-CoV-2 virus, which affects a person's lungs making their respiratory system so bad that it can lead to death. In this project, the aim is to predict the evolution of patients suffering from COVID-19 from X-ray images. These images have been obtained through the dataset that HM Hospitales group made public, a group of hospitals presents and operating in Spain. The images used come directly from the source, and have not been treated in any way to safeguard their veracity and to make them as close as possible to reality. In this project, we have taken into consideration two circumstances that can be observed today. First, the lack of data and datasets corresponding to people suffering from Covid-19, and second, the lack of projects that only take into account these people in order to predict the evolution of the disease. Due to the problem that this virus has caused for the whole world, this project attempts to predict, through an X-ray image of the patient's rib cage, whether he will have a serious evolution, which could even lead to death, or if, on the contrary, he will have a milder evolution and not require admission to the ICU. For the realization of these predictions, two different approaches have been developed, one using only transfer learning and the other with images generated through DCGANetworks. The results obtained in the first approach were not entirely satisfactory: • Model implemented with ResNet50: Validation accuracy 76.312%, Test accuracy 52%, False negative errors 55.20%. • Model implemented with VGG16: Validation accuracy 74.722%, Test accuracy 51%, False negative errors 59.92%. The results in the second approach that includes the images artificially generated by DCGAN within the dataset significantly improved the results of the first approach at the level of test accuracy and False negative errors (test result which wrongly indicates that a condition does not hold): • Model implemented with ResNet50: Validation accuracy 73.450%, Test accuracy 53%, False negative errors 51.47%. • Model implemented with VGG16: Validation accuracy 59.744%, Test accuracy 60%, False negative errors 26.72%. As can be seen, the DCGAN network has helped to improve the results obtained in the approach where it was not applied, reducing more than 50% of false negative errors, the most problematic in medicine. The results have been satisfactory, although the results can continue to be improved in order to have a real application. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|