Article Preview
TopIntroduction
The COVID-19 pandemic has become a serious health crisis globally since December 2019 (Zhu et al., 2020; Benvenuto et al., 2020; Wang et al., 2020; Ugail et al., 2021) The symptoms of COVID-19 are fever, dry cough, tiredness, diarrhea, headache, loss of taste and smell. The serious symptoms include shortness of breath as well as an inability to walk or stay awake and may require emergency medical care (Huang et al., 2020; Ng et al., 2020). The exponentially increasing COVID-19 positive cases have limited the accurate and early detection of disease in patients due to a shortage of testing kits. One of the methods of diagnosing the disease is by segmenting the infections in the lung CT scan. CT images play a crucial role in detecting COVID-19 by segmenting the lesions from the images (Lei et al., 2020; Li et al., 2020). Manual segmentation is laborious and time-consuming and is also subjected to variations in annotations by different radiologists. These challenges can be alleviated by automating the process of infection segmentation from the CT scans (shan et al., 2020). Deep learning techniques have been frequently used in the medical field in recent times and these techniques have proved to be successful (Wu et al., 2020; Huang et al., 2020; Jha et al., 2020; Patro et al., 2020; Khan et al., 2020; Nandhini Abirami et al., 2021). Deep learning techniques can be adopted to accurately identify and segment the infections from the CT images (Khan et al., 2021; Rauf et al., 2021).
Infection segmentation from CT scans is difficult due to irregularity in the shapes of the lesion and low contrast between normal and infected lesions. Ground glass opacities were observed in the early stages of infection, while pulmonary consolidation was observed in the later stages (Ye et al., 2020; Akbari et al., 2020). In recent times, deep learning models have frequently been used for accelerating the diagnosis of COVID-19 via chest X-ray images and CT scans (Mangal et al., 2020; Shi et al., 2020; Ma et al., 2020). The performance of the best CNN models in classifying COVID-19 is evaluated in (Apostolopoulos et al., 2020). For distinguishing between people with infected and normal lungs, a deep learning model was employed in (Toğaçar et al., 2020). A multi-objective differential evolution (MODE) and CNN were used to classify COVID-19 infections using chest CT images (Singh et al., 2020). A semi-supervised deep learning model was presented in (Zheng et al., 2020) to detect COVID-19 from CT images (Zheng et al., 2020). An anomaly detection model was proposed to differentiate among pneumonia classes using COVID-19 chest X-ray images (Zhang et al., 2020).
Classic lung CT segmentation techniques are computationally expensive, fail to handle attenuation variation, and fail to segment regions with high levels of abnormality (Roy et al., 2020). Also, the segmentation of infections from lung CT images is challenging, and its performance is affected by (i) lung inhomogeneity, (ii) image acquisition protocol, (iii) scanner type, and (iv) GGO (Ground Glass Opacity) nodules. Segmenting GGO nodules is challenging because of subtle boundaries that are hard to delineate. Also, the variations in the position and size of the infections in the CT images poses challenges in segmenting the lesions from CT slices.