Article Preview
TopI.Introduction
The ongoing COVID-19 pandemic is finally reaching all the continents on earth. The epidemic began in November 2019 in China. As of July 04, 2020, over 180M positive cases, causing around 40 lakhs deaths (according to worldmeters.info/coronavirus/) worldwide. Three COVID-19 vaccines have been approved by national regulatory authorities and are available for use from December 2020 (WHO, 2020) . Until WHO approves vaccines for regulation and the entire population will be vaccinated, the most reliable and widely used testing method for detecting coronavirus is RT-PCR (Yang X. et al., 2020). The major problem with RT-PCR is getting the result, which requires 4-6 hours to generate the results (Mishra AK et al., 2020). As this virus is spread through the droplet, a delayed outcome is a major concern because an infected person can contact others and infect them as well. Therefore, there is a need for alternative methods that hospitals and physicians can use to diagnose COVID-19 (Yang X. et al., 2020). Medical images such as CT scans and X-rays help to diagnose COVID-19 with clinical symptoms that suggest lung infection (R.K. Gupta, 2021; Cai X., 2020). It is demonstrated that CT scan tests exhibit higher sensitivity than the RT-PCR ones. This point is further validated by the sensitivity values in which CT scans and RT-PCR tests obtained 98% and 71%, respectively. However, the diagnosis period of CT scans continues to be a significant restriction; even experienced radiologists require approximately 21.5 minutes to review the test data of individual cases (Huang Z., 2020).To combat the COVID 19 epidemic, many computers vision and artificial intelligence approaches and procedures have been developed, along with methods of segmentation and classification (Ouyang W. et al., 2019). These methods can be divided into two classes: standard Machine Learning (ML) and Deep Learning (DL) (Wang S. et al., (2017). DL is a bio-inspired approach in which the AI model mimics the human brain in such a way to process data and use that knowledge to detect objects for decision-making, speech recognition, and language translation. Image segmentation assists medical researchers and clinical practitioners in many ways and played a significant role in medical diagnosis and clinical decision-making. Segmentation is the separation of Regions Of Interest (ROIs) from the whole object or region and can be used to carry out a quantitative assessment for further diagnostic. Homogeneity in the intensity of pixels, undesirable artifacts, and proximity in the grey level of distinct tissues of a body part are the main obstacles confronting segmentation methods (Oulefki A., 2020).The methods for segmentation can be classified into three classes: manual, semi-automatic, and automatic (M.Choudhary et al. 2019; 2020). The monotonous and complexity of manual segmentation methods are ubiquitous and influenced by inter and intra-observation variability. Semi-automatic approaches have been widely utilized and integrated with open-source software packages. In contrast, automatic procedures are free of user intervention. All of these approaches have their own set of benefits and drawbacks. Researchers, however, still striving to simplify the automatic segmentation process but still confronting the challenges of the segmentation process (Yuan X., 2010), like the lack of a generalized solution that can be applied to an increasing number of different areas of interest, variations in different ROI attributes and variation in the noise of each object or signal of homogeneity associated with the images (Elnakib A., 2011).(Shi et al., 2020) emphasized that numerous AI systems were proposed to support COVID-19 diagnosis, but few studies connected with CT scan segmentation. However, when we have a short dataset, these strategies are not efficient and trustworthy. Detecting infection in CTs by experienced doctors with naked eyes is tedious, there are chances of missing out on small ROI. Therefore, there is a need for an automated method that can automatically segment the infection region. Different segmentation approaches are extensively utilized to segment ROIs in CT. Some typical segmentation methods include thresholds, active contours, differential operators, watershed, and regional growth. The threshold methods have a significant false-negative rate for semi-transparent nodules. The active contour is performed on average. The Watershed algorithm has a high false-negative rate in the vasculature in solitary nodules and nodules. Due to the requirement of a highly sensitive filter for removing noise, a different operator method is not suitable.To address all the above issues, this article proposes Gaussian Mixture Model-based Hidden Markov Random Field (GMM-HMRF) for segmentation of lung CTs of COVID-19 patients and normal lung CT images. Then the obtained segmented images are fed to a modified Resnet-18 CNN model for training and classification tasks. Resnet-18 is modified by adding some layers of increased filter size. The proposed method successfully segmented the ROIs in lung CTs and the segmented dataset helped get higher accuracy during Resnet 18.The main contribution of this paper is as follows:
- 1.
A novel segmentation approach based on GMM-HMRF.
- 2.
Generalizing the GMM-HMRF trained using expectation maximization (EM)
- 3.
Modified ResNet18 deep model by enhancing the architecture including extra layers.
The rest of the paper is organized as follows. In Section 2, the existing work on the segmentation of medical images and classification using DL methods is discussed. In Section 3, The steps related to the proposed algorithm and methods are discussed. Section 4 contains the experimental details and the results obtained. In the last section, we concluded our study.