The COVID-19 pandemic is one of the most significant threats to the general population's health in the 21st century. In this study, a novel meta-learning based FSS model is proposed. This model is realized as an adaptive relation network built on Deeplabv3+ for training the support sets and a convolutional network with swish activations functions for non-linear metric learning. The performance of this model that was trained using supervised and semi-supervised learning algorithms on two public datasets is significantly better. This model obtains a global accuracy of 0.8396 for ground glass opacity (GGO) and consolidation segmentation and 0.9996 for entire lung infection segmentations correspondingly. In addition, the model that was proposed generalizes well with data that has not yet been seen and has the potential to be expanded to the identification of other infections in image volumes that are rendered in three dimensions and four dimensions.
TopIntroduction
The most recent form that COVID-19 has taken is a major global public health emergency in its own right. Polymerase Chain Response (PCR) tests for initial screening are not readily available and found to be negative in cases with mild disease. Further, this test also reports false positives due to viral contamination in reagents and primers. Although other diagnostic tests have been used to investigate for potential presence of COVID-19, such as antibody tests, their efficacy has not been proven and these tests require specialized and expensive equipment. The significance of imaging examinations for COVID-19 management has been recognized by the healthcare communities.
Initially, chest X-rays (Pereira, 2020) were preferred for COVID-19 screening owing their easy availability and affordability. However, sensitivity and specificity of this modality are low in diagnosing COVID- 19 patients with milder disease. An investigation by Dong et al. (2020) has shown that typical image features and subsequent changes in them are significant in COVID-19 management and advocates the usage of Computed Tomography (CT) to progress the alacrity and precision of diagnosis.
In this line, Simpson et al. (2020) predicted the possible usage of CT scanning in COVID-19 management and proposed four categories of reporting the findings relevant to COVID-19. These authors report the presence of small discrete pulmonary nodules, atypical of COVID-19. Generally, pulmonary nodules are evidenced in Chest x-ray and CT examinations, manifesting as circular or oval shaped growths, caused by Lung cancer, Carcinoid, Lymphoma and Metastatic tumors. Wide prevalence of lung cancer has provoked extensive research and development of automated mechanisms for pulmonary nodule detection (Pena, 2016; El-Regaily, 2018; Zhang, 2019) from CT scans.
Recently, few investigations have reported the presentation of COVID-19 as solitary lung nodules. A case report by Rasekhi et al. (2020) suggests that patients with pulmonary nodule must be examined for COVID-19. Similarly, Ahmed and Rice (2020) also report a solitary nodule in the chest X- ray of COVID-19 patients. A study by Xia et al. (2020) also divulges the possibility of a sub-center Ground Glass Nodule (GGN) in the CT images of COVID-19 infected patients. These findings reinstate the significance of lung nodule detection in CT images for lung cancer and COVID-19 management.
Conventional automated systems for lung nodule detection are deployed as classification models trained with hand-crafted (Pu, 2008; Shi, 2016) features such as size, color, texture and shape, extracted from CT images. However, extraction of these features is time consuming and they suffer from lack of normalization and uniformity.
The pioneering FSS approach proposed in Mondal (2018) demonstrates the efficacy of Generative Adversarial Networks for brain segmentation with labeled and unlabeled multi-modal 3D MRI images.
In COVID19 detection, availability of annotated CT datasets is constrained by time-consuming image acquisition procedures and lack of experts for image labeling. Owing to this limitation, recently FSL based models are being explored by different researchers to build classifier models for COVID19 detection with limited data. Fan et al. (2020) have proposed Inf-Net, a lung infection segmentation network for COVID19 detection employing Reverse Attention (RA) and Paralleled Partial Decoder (PPD) mechanisms. The authors have also built a semi-supervised version of this model to address the lack of labeled data, employing random sampling strategy on unlabeled chest CT images to enlarge the dataset.
Generally, FSS is characterized as a meta-learning problem in which, a meta-learner trained with several segmentation tasks trains a segmentation network. However, some of the existing works on FSS based on Siamese networks (Rakelly, 2018; Hu, 2019; Shelhamer, 2018) are not suitable for multi-class segmentations as they limit the number of segmentation classes to two.
In this paper, it addressed a novel FSS network called CT-FSS, for multi-class semantic segmentation of lung infections from chest CT images for COVID19 management. The meta-learner of this network is realized as a relation (Chen, 2018) network, which non-linearly learns prior knowledge from DeepLabv3+ (Sung, 2018) networks trained with segmentation tasks on different support sets.