KSU - Covid19 Challenge

In the current critical situation of COVID-19 spread throughout the world appropriate image assessment can help to optimize treatment for patients admitted to hospitals. Currently imaging by x-ray radiography is standard, in uncertain cases with CT. Patterns typical for COVID-19 commence with predominantly peripheral ground glass opacities visible on CT, followed by interstitial changes and consolidations that can become extensive at later disease stage, associated with a poor prognosis. Radiographs are less sensitive and specific compare to CT but still contain valuable information (W. Liang et al., JAMA 2020). Imaging may help improve patient stratification, e.g. predicting a poor outcome.

Artificial intelligence in imaging

Classification of patients based on AI analysis of chest radiographs has been documented for lung disorders including tuberculosis, pneumonia,... (I. Sirazitdinov et al. Comput Electr Eng 2019). AI methods may be helpful in assisting the radiologist, pointing to suspicious image features. AI method development require large datasets and technical expertise. However, application later can be done for single patients and on patients groups. Currently, the number of COVID-19 cases in public repositories is limited. We have extensively searched the web but the number of cases made available is still limited. How to deal with this situation to come up with a powerful AI approach?

Approaches tried

Used transfer-learning in Tensorflow. Experimented with VGG16, InceptionV3 and Xception model trained on ImageNet.

Model with pre-trained inceptionV3 model performed slightly better than VGG on pre-COVID dataset(85% accuracy), whereas pre-trained VGG model performed significantly better on COVID dataset(72% accuracy). Slight augmentation of the training data(rotation, brightness change) was alos performed. Xception and ensemble model couldn't be completely tested due to training time constraints.

Built With

• python
• tensorflow