Evaluation of EfficientNet models for COVID-19 detection using lung parenchyma

• Published in: Neural computing & applications Vol. 35; no. 16; pp. 12121 - 12132
• Main Authors: Kurt, Zuhal, Işık, Şahin, Kaya, Zeynep, Anagün, Yıldıray, Koca, Nizameddin, Çiçek, Sümeyye
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2023; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Computational Biology/Bioinformatics; Computational Science and Engineering; Computed tomography; Computer Science; Coronaviruses; COVID-19; Data Mining and Knowledge Discovery; Datasets; Image Processing and Computer Vision; Image segmentation; Lungs; Original; Original Article; Probability and Statistics in Computer Science; Deep learning; means; EfficientNet; COVID-19 detection; CT scan; Lung parenchyma; K-means
• ISSN: 0941-0643; 1433-3058; 1433-3058
• DOI: 10.1007/s00521-023-08344-z

When the COVID-19 pandemic broke out in the beginning of 2020, it became crucial to enhance early diagnosis with efficient means to reduce dangers and future spread of the viruses as soon as possible. Finding effective treatments and lowering mortality rates is now more important than ever. Scanning with a computer tomography (CT) scanner is a helpful method for detecting COVID-19 in this regard. The present paper, as such, is an attempt to contribute to this process by generating an open-source, CT-based image dataset. This dataset contains the CT scans of lung parenchyma regions of 180 COVID-19-positive and 86 COVID-19-negative patients taken at the Bursa Yuksek Ihtisas Training and Research Hospital. The experimental studies show that the modified EfficientNet-ap-nish method uses this dataset effectively for diagnostic purposes. Firstly, a smart segmentation mechanism based on the k -means algorithm is applied to this dataset as a preprocessing stage. Then, performance pretrained models are analyzed using different CNN architectures and with our Nish activation function. The statistical rates are obtained by the various EfficientNet models and the highest detection score is obtained with the EfficientNet-B4-ap-nish version, which provides a 97.93% accuracy rate and a 97.33% F 1-score. The implications of the proposed method are immense both for present-day applications and future developments.