A mobilenet based CNN model with a novel fine tuning mechanism for COVID-19 infection detection

A new coronavirus called COVID-19, which emerged in Wuhan, China, in December 2019, is very dangerous because of its rapid global spread worldwide, causing severe acute respiratory syndrome. It has infected 349.641.119 people worldwide and caused over 5.592.266 deaths, as reported in January 2022. Although there are several different diagnostic methods, polymerase chain reaction (PCR) is considered the gold standard for laboratory diagnosis of the COVID-19 pathogen. However, test results are obtained within a few hours to two days, and this relatively late response is the main barrier to early intervention. Researchers have focused on alternative methods that use x-ray imaging to shorten the time to diagnose the disease. This study proposes a deep-transfer learning approach with novel fine-tuning mechanisms for detecting COVID-19 disease using chest X-ray images. The model is based on the MobileNetV2 architecture. To evaluate the proposed model, we used a combined dataset from two publicly available databases containing three classes: normal, COVID-19, and pneumonia X-ray images. In our approach, we proposed one classical and two new fine-tuning mechanisms to increase classification accuracy and achieved average accuracy rates of 95.62%, 96.10%, and 97.61% for 3-class cases with five-fold cross-validation. In addition, our third model reduced 81.92% of the total fine-tuning operations and achieved better results. The numerical results show that the proposed approach achieves promising results when raw data without complex preprocessing steps are used.