Enhancement the Knowledge Distillation for COVID-19 Detection from Computed Tomography Images by Using Neural Architecture Search

Over the past few years, the spread of COVID-19 has significantly impacted people’s lives globally, necessitating rapid diagnosis to interrupt its spread and reduce infection risk. CT imaging has been recognized as a vital tool for detailed insights into the disease’s manifestations, with deep learning proving highly effective in improving diagnostic accuracy and efficiency. However, despite achieving high accuracy, previous works have faced challenges related to high computational costs, reliance on scarce labeled data, and difficulty generalizing to diverse datasets. To address these issues, a novel model to enhance knowledge distillation for COVID-19 detection in CT images is proposed. The contributions include enhancing Knowledge Distillation (KD) using Neural Architecture Search (NAS), implementing NAS to decrease false positive rates, and introducing a model for detecting COVID-19, aiding faster and more efficient clinical decisions. The algorithm’s ability to process previously unseen medical images is highlighted, with NASNetLarge as the teacher model and NASNetMobile as the student model demonstrating high performance and satisfactory classification results. The proposed method achieves high accuracy, reduces false positives, and enhances overall diagnostic performance. The teacher model demonstrates satisfactory performance in detecting CT images, with a high accuracy of 99.02%. In addition, the student model, which performs as a lightweight model, shows good classification performance with an accuracy of 98.77%.