Designing Effective Models for COVID-19 Diagnosis through Transfer Learning and Interlayer Visualization

Abstract

Creating a model from scratch that fits the dataset can be laborious and time-consuming. The level of difficulty in designing a new model can vary depending on factors such as the complexity of the model and the size and characteristics of the dataset. Factors such as the number of variables in the dataset, the structure of the data, class imbalance, and the size of the dataset are important in deciding which model to use. In addition, long experimental studies are required to design the most appropriate model for the dataset. In this study, we investigated how transfer learning models can be utilized to solve this problem. Experimental studies were conducted on the Covid-19 dataset with transfer learning models and the most successful transfer learning models were identified. Then, layers that did not contribute to the performance of the transfer learning models and could not extract the necessary features from the dataset were identified and removed from the model. After removing the unnecessary layers from the model, new models with fast, less complex and fewer parameters were obtained. In the studies conducted with the new models derived from the most successful transfer learning models with the inter-layer imaging method, the classes were classified with an accuracy of %98.8 and the images belonging to the Covid-19 class were classified with a precision of %99.7.