AU2021103578A4 - A Novel Method COVID -19 infection using Deep Learning Based System - Google Patents

Abstract

Coronavirus illness has infected millions of individuals globally and is contaminating them at an alarming rate, putting a strain on the health system's capabilities. PCR screening is the analytic technique of choice for COVID-19 discovery. CT imaging has demonstrated its diagnostic competence in asymptotic patients, establishing it as a reliable diagnostic support tool for COVID-19. The high prevalence of COVID-19 contaminations in CT slices enables the tracking of illness progression using automated contamination segmentation techniques. Nevertheless, COVID-19 infection regions exhibit a great degree of heterogeneity in terms of scale, form, contrast, and concentration, posing a substantial challenge to the segmentation approach. The present invention provides an automated segmentation technique based on deep learning for the discovery and delineation of COVID-19 infections in CT images. Additionally, as compared to existing techniques, this suggested innovation will need less time and money to identify the infection and annotate the infection regions. This innovation will assist physicians in determining the course of covid-19 illness, as well as quantifying the disease's load and severity, by segmenting the lung organ from the CT scan as a region of concern and then segmenting the infections contained inside it.

Description

TITLE OF THE INVENTION

A novel method COVID -19 infection using Deep Learning Based System

FIELD OF THE INVENTION

The current disclosure is mainly linked to a Deep Learning Based System for Detection of COVID-19 Infections in Humans, which will aid doctors in determining the course of the covid-19 sickness, as well as measuring the burden and seriousness of the condition.

BACKGROUND OF THE INVENTION

Coronavirus illness has infected millions of individuals globally and is contaminating them at an alarming rate, putting a strain on the healthiness system's capabilities. PCR showing is the diagnostic technique of choice for COVID-19 recognition. CT imaging has demonstrated its diagnostic competence in asymptotic patients, establishing it as a reliable diagnostic support tool for COVID-19. The high prevalence of COVID-19 infections in CT slices enables the tracking of illness progression using automated infection dissection techniques. However, COVID-19 infection regions exhibit a great degree of heterogeneity in terms of scale, shape, contrast, and intensity, posing a substantial challenge to the segmentation approach. The present invention provides an automated segmentation technique based on deep learning for the detection and delineation of COVID-19 infections in CT images. The major goal of this invention is to present a method for detecting covid-19 infection in order to assist physicians in determining the course of covid-19 disease, as well as to quantify the disease's burden and severity.

SUMMARY OF THE INVENTION

The new coronavirus illness (COVID-19) pandemic has spread to 213 nations and territories, with two transmissions occurring outside. Based on the current COVID-19 dashboard at Johns Hopkins University's Center for Systems Science and Engineering (CSSE), more than 19,172,505 instances of COVID-19 have been described, including 716,327 fatalities. The high prevalence of COVID-19 infection accounts for the fast increase in reported cases. COVID-19 causes significant respiratory difficulties, such as fever, coughing, and fatigue, as well as a variety of other symptoms. These symptoms, in particular, can progress to serious pneumonia in patients with compromised immune systems. The reverse transcription polymerase chain reaction (RT-PCR) test is the most often used method for detecting viral RNA using a nasopharyngeal swab. Numerous investigations, however, have demonstrated that the RT-PCR test has a high percentage of false negatives, necessitating repeated testing for proper diagnosis. Additionally, the RT-PCR test is limited in availability owing to the shortage of manufacturing material, and the testing method is time intensive, preventing quick and reliable screening. Computed tomography (CT) is used in conjunction with real time polymerase chain reaction (RT-PCR) for COVID-19 screening, as a significant proportion of CT scans were acquired from infected individuals.

Thus, these regions of COVID-19 infection can be detected in asymptomatic individuals by evaluating ground glass opacity (GGO) and indications of lung consolidation that may develop at various phases of the illness. Visual CT imaging analysis may aid in the identification of COVID-19 illness by providing techniques for classifying prevalent infection patterns such as ground glass opacity (GGO) and pulmonary consolidations. Systems in this field perform three distinct image processing tasks. Second, CT categorization, in which the patient is classified as having the illness or not having it. Second, illness infection identification, with bounding boxes highlighting infection regions. Segmentation of the infection area and quantification of disease load by pixel level classification is the third job. Manual segmentation of COVID-19 infections is a lengthy and time-consuming operation. Additionally, it is very dependent on the talents of the physician or doctor performing the segmentation process. The purpose of the general FCN is to categorise pictures, where an image is input and one mark is output.

However, in the COVID-19 chest CT examination, in addition to categorization, the field of abnormality must be localised and segmented. Researchers began utilising FCN for COVID-19 illness to aid physicians and radiologists in diagnostic and prognostic activities, which led in increased accuracy and decreased inspection time. Numerous researchers have proposed the use of deep learning systems to aid in the fight against the increased spread of COVID-19 illness. The majority of recently suggested deep learning methods are based on the identification (classification) of patients infected with COVID-19 illness by CT screening, which is possible owing to the widespread availability of CT scans and does not need extremely uncommon radiological annotations. COVID-Net, as described in the literature, is a deep neural network optimised for identifying COVID 19 instances from chest X-ray pictures that are open source and accessible to the general public. It serves as an example of the categorization of COVID-19 diagnostic systems. Using pulmonary CT scans, another study suggested a 3D deep learning system capable of differentiating COVID-19 pneumonia from Influenza-A viral pneumonia and healthy patients. To identify COVID-19, another software system based on weakly-supervised deep learning was created utilising 3D CT volumes. Numerous deep learning algorithms have been proposed to aid in the diagnosis of COVID-19 in clinical practise, however only a handful are connected to the delineation of infection from CT images. The majority of these recommended approaches rely on the installation of U-net FCN as its backbone. Other works have proposed their own deep COVID-19-oriented networks. For the majority of proposed methods, the common issue is insufficiently labelled CT images for deep network training that cannot be retrieved quickly. Because the procedure of marking infection sites is time consuming, expensive, and reliant on the radiologist's skill. This invention offers an automated segmentation technique based on deep learning for the purpose of detecting and delineating COVID-19 infections in CT images. The approach begins by segmenting the lung organ from the CT image as a potential source of infection, followed by segmenting the infections contained inside it. The approach enables physicians to track the course of the disease, as well as to quantify the disease's burden and severity.

DETAILED DESCRIPTION OF THE INVENTION

The new coronavirus illness (COVID-19) pandemic has spread to 213 nations and territories, with two transmissions occurring outside. According to the COVID-19 dashboard at Johns Hopkins University's Center for Systems Science and Engineering (CSSE), more than 19,172,505 instances of COVID-19 have been reported (the number is rising), including 716,327 fatalities. The high prevalence of COVID-19 infection accounts for the fast increase in reported cases. COVID-19 causes significant respiratory difficulties, such as fever, coughing, and fatigue, as well as a variety of other symptoms. These symptoms, in particular, can progress to serious pneumonia in patients with compromised immune systems. The reverse transcription polymerase chain reaction (RT-PCR) test is the most often used method for detecting viral RNA using a nasopharyngeal swab. Numerous investigations, however, have demonstrated that the RT-PCR test has a high percentage of false negatives, necessitating repeated testing for proper diagnosis. Additionally, the RT-PCR test is limited in availability owing to the shortage of manufacturing material, and the testing method is time intensive, preventing quick and reliable screening. Computed tomography (CT) is used in conjunction with real time polymerase chain reaction (RT-PCR) for COVID-19 screening, as a significant proportion of CT scans were acquired from infected individuals. CT scanning diagnostic investigations on individuals with COVID-19 illness show that regions of infection may emerge on CT scans prior to the onset of disease symptoms. Thus, these regions of COVID-19 infection can be detected in asymptomatic individuals by evaluating ground glass opacity (GGO) and indications of lung consolidation that may develop at various phases of the illness. Visual CT imaging analysis may aid in the identification of COVID-19 illness by providing techniques for classifying prevalent infection patterns such as ground glass opacity (GGO) and pulmonary consolidations. Systems in this field perform three distinct image processing tasks. Second, CT categorization, in which the patient is classified as having the illness or not having it. Second, illness infection identification, with bounding boxes highlighting infection regions. Segmentation of the infection area and quantification of disease load by pixel-level classification is the third job. Manual segmentation of COVID-19 infections is a lengthy and time-consuming operation. The purpose of the general FCN is to categorise pictures, where an image is input and one mark is output. However, in the COVID-19 chest CT examination, in addition to categorization, the field of abnormality must be localised and segmented. Researchers began utilising FCN for COVID-19 illness to aid physicians and radiologists in diagnostic and prognostic activities, which led in increased accuracy and decreased inspection time. Numerous researchers have proposed the use of deep learning systems to aid in the fight against the increased spread of COVID-19 illness. The majority of recently suggested deep learning methods are based on the identification (classification) of patients infected with COVID-19 illness by CT screening, which is possible owing to the widespread availability of CT scans and does not need extremely uncommon radiological annotations. COVID-Net, as described in the literature, is a deep neural network optimised for identifying COVID 19 instances from chest X-ray pictures that is open source and accessible to the general public as an example of the categorization of COVID-19 diagnostic systems. Using pulmonary CT scans, another study suggested a 3D deep learning system capable of differentiating COVID-19 pneumonia from Influenza-A viral pneumonia and healthy patients. To identify COVID-19, another software system based on weakly-supervised deep learning was created utilising 3D CT volumes.

Numerous deep learning algorithms have been proposed to aid in the diagnosis of COVID-19 in clinical practise, however only a handful are connected to the delineation of infection from CT images. The majority of these recommended approaches rely on the installation of U-net FCN as its backbone. Other works have proposed their own deep COVID-19-oriented networks. For the majority of proposed methods, the common issue is insufficiently labelled CT images for deep network training that cannot be retrieved quickly. Because the procedure of marking infection sites is time consuming, expensive, and reliant on the radiologist's skill. This invention offers an automated segmentation technique based on deep learning for the purpose of detecting and delineating COVID-19 infections in CT images. The approach begins by segmenting the lung organ from the CT image as a potential source of infection, followed by segmenting the infections contained inside it. The approach enables physicians to track the course of the disease, as well as to quantify the disease's burden and severity. framework that is proposed: The suggested segmentation method's flowchart is depicted in Fig. 1, and it comprises of two major stages that are carried out consecutively. The initial stage is lung segmentation from standard chest CT slices, followed by COVID-19 infection. Open Access COVID-19 Datasets The proposed FCN networks in this invention are trained and tested using COVID-19 chest CT datasets obtained from a number of sources. The requisite training classes were identified using all CT datasets: history, lung organ, and COVID-19 infection. The first COVID-19 dataset will be made publicly available by the Italian Society of Medical and Interventional Radiology (SIRM). The COVID-19 CT segmentation dataset consists of 100 axial slices from over 40 distinct COVID-19-infected CT pictures, each having radiologist-segmented lung infections. The second collection of public datasets comes from Radiopedia and consists of nine COVID-19 chest volumetric CT scans with accompanying ground-truth. A radiologist diagnoses about 373 slices from each of the nine datasets as positive and defined. Ma et al. are compiling and presenting a third recently published public dataset, which includes 20 annotated COVID-19 chest CT images, ten Coronacases Initiative CT scans, and ten more Radiopedia CT scans. These CT scans are publicly accessible under a Creative Commons BY-NC-SA licence and have been labelled by two radiologists and verified by an expert radiologist on all twenty COVID-19 CT images. Enhancement of the appearance of COVID-19 infections: Segmentation of infection areas within the lung ROI presents a number of difficulties, including the presence of fuzzy edges and a high degree of inhomogeneity within infection areas. The COVID 19 infection area is faintly contrasted on chest CT scans; the surrounding tissues lack a distinct border. Additionally, infection areas in CT slices exhibit a high degree of variability in terms of texture, size, and location. To increase the identification of infection regions, the segmented lung (ROI) from the CT image is enhanced using tensor-based Edge Enhancing Diffusion (EED) filtering. EED filtering makes use of a diffusion tensor to adjust the diffusion to the picture structure. EED filters aid in increasing contrast, filtering noise to improve strength uniformity, and preserving form boundaries. EED filtering is used to enhance the contrast of infection regions in order to facilitate the detection and segmentation of COVID-19 infection sites by increasing the uniformity of severity within these areas while maintaining lung parenchyma limitations. This approach attempts to improve the process of FCN training by isolating and learning the essential characteristics that differentiate infection sites from surrounding tissues. Extraction of Training Patches: The distribution and position of infection regions within CT scans are unclear, which is a key issue of the proposed deep learning FCN for improved training. Additionally, if the CT slice contains regions of infection, the distribution of infection within the slice is significantly skewed, as the infection may comprise just a tiny proportion of the slice. Thus, utilising the complete CT slices as training patches may result in a severe bias against the context, which is a well known issue of semantic segmentation in medical imaging. To address this issue, training slices are generated from infection-containing slices using random patch sizes. The training datasets consist of lung slices generated solely from the lung ROI (not the whole CT slice). Due to the fact that the lung organ appears in numerous slices of varying sizes, the patches excised have varying resolutions. Two cascaded deep FCNs are coupled successively in this study to partition the COVID-19 lung organ and subsequently the infection regions. A modification to the U-net design, which consists of five layers and serves as the backbone of the planned FCN network. A U-network is made up of two paths: one for encoding and one for decoding. At each stage of the encoding process, three operations are performed: convolution, activation function (ReLU), and batch normalisation. These procedures are performed twice sequentially in each level block, followed by a max-pooling operation before proceeding to the next point. The kernel size is 3x3 for convolutions and 2x2 for max-pooling. After each level, the resolution of the feature is halved. The network decoding approach restores the original input size by following the same operation sequence conyy, ReLU, BN) but substituting upsampling for max-pooling at each step. Additionally, the appropriate function is concatenated to each decoding stage's input from the encoding path. The last stage of the decoding route is a 1x convolution with a sigmoid activation function to categorise the feature map and generate the final binary prediction map using the dice coefficient metric. While the growth in network depth is unavoidable in FCN, it introduces the problem of the vanishing gradient, as additional layers are heaped on top of each other and gradient information is washed out, slowing down training and degrading efficiency. Numerous deep network topologies have been proposed to address this problem, however DensNet and ResNet are often recognised as advances in terms of performance. Each layer is connected to all subsequent layers in DensNet, where feature maps produced using various filter sizes are concatenated from previous layers, significantly thickening the model as channels are merged following each convolution operation. In ResNet, on the other hand, this does not occur since the addition operation is utilised to merge the prior input identity with the output function map. Within the

ResNet block, a shortcut (skip link) from the block input (identity) links to the block output feature, bypassing the stacked layers. DensNets are also known as memory hungry networks, as back-propagation enables the storage of the whole layer outputs, which consumes more memory and operates slowly. On the other side, ResNet's suggestion is to include tensors, however it has been claimed that adding function mappings directly affects the gradient flow across the network, as it sums the values of the features. As a result, the concatenation operation is preferred, as it preserves the function maps, whereas the summing procedure corrupts the feature maps for both the convolution operation and the source of the skip connections. ResDense is a critical component of the proposed network design. In the proposed ResDense, dense ties (concatenation) are employed between residual blocks rather than convolution layers. In terms of flow and memory feature maps, the suggested ResDense block suitably refines feature values via the use of residual blocks and occasionally memorises the refined feature values through the use of dense connections between residual blocks. The encoder and decoder pathways for the proposed FCN network are dependent on ResDense blocks in this study. Each level in the proposed network's contracting and extending pathways is created using the ResDense block. As a result, at the conclusion of each stage of the encoding route, the feature map's depth is doubled and concatenated with the block input. Resized annotated 2D slices with a patch size of (256 256) are used to train the proposed network. Using zero mean and unit variance normalisation, all CT slices were patch-wise corrected. Keras1 and the TensorFlow backend are used to implement the ResDense FCN architecture. To split the lung and then the diseased region sequentially, two ResDense networks were constructed. The Adam optimizer with a learning rate of 0.0001 is used to train the network and modify the parameters. For lung segmentation and infection segmentation, both networks are trained with 20 epochs and batch sizes of 32 and 16, respectively. To update the model parameters and track the convergence of the network training, the soft dice coefficient loss was utilised. In network design, the Batch Normalization

(BN) layer is used to prevent erroneous increases or decreases in produced data between network levels. The pixel-wise sigmoid activation feature is used in the network's last layer. Many efforts have been taken to increase the certified network's efficiency as a result of the uneven distribution of lung tissue classes and COVID-19 infections. To begin, the soft dice loss measure is utilised throughout the training phase to determine the overlap between the ground-truth patches and the area within the lung ROI designated as infected by the network. Additionally, the second FCN network is trained exclusively within the lung ROI to identify characteristics that distinguish COVID-19 infections from lung tissues. After cutting the lung region, the training patches are retrieved only from the CT slice. Additionally, it is guaranteed that the training patches utilised contain the appropriate mask and any patches without an annotation mask are removed from the training phase. Performance Measures: A battery of performance tests is performed to evaluate the proposed technique quantitatively by segmenting COVID-19 infection and the lung organ from CT images. To begin, the Dice coefficient (DSC) is a measure of overlap between the successfully segmented class and the ground truth in terms of the average scale. The dice metric (DSC) is defined as DSC = (2*TP)/(2*TP+FN+FP), where TP (true positive) and TN (true negative) are the number of voxels accurately categorised in the segmented class and context, respectively. The terms FN (false negative) and FP (false positive) refer to the number of voxels that are neither contextually relevant nor accurately categorised as segmented. The second metric is sensitivity, which is defined as the ratio of correctly segmented voxels (TP) to ground truth voxels (TP+FN). Sensitivity = TP/(TP+FN). The sensitivity metric indicates the system's ability to separate the desired class voxels properly. The accuracy of the ratio of correctly segmented non-class voxels (TN) to the total number of non-class voxels measured is the third metric. Specificity =

TN/(TN+FP) is used to demonstrate that the technique is capable of segmenting voxels that do not belong to the desired class.

This invention offers a deep learning framework for COVID-19 lung infection segmentation on chest CT images. The created FCN employed a U-net architecture as a backbone for each level of the encoding and decoding pathways, utilising proposed ResDense blocks. Due to the concatenation skip relation in each ResDense block, the feature maps of the infection regions and lung history pass through the network without significantly changing their values, which increased network learning and segmentation efficiency. Additionally, the technique incorporates an EED phase that enhances the appearance of infection regions in CT slices by increasing their contrast and strength uniformity. The qualitative and quantitative evaluation findings demonstrate the system's efficacy and capability for segmenting COVID-19 infection regions from CT images. The system is trained and validated using a variety of datasets from a variety of sources, demonstrating its generalizability and making it one of the most promising methods for automated COVID-19 infection diagnosis and clinical routine analysis.

Claims (5)

CLAIMS: We Claim:

1. We assert that the given invention offers a deep learning-based framework for COVID-19 lung infection segmentation in chest CT scans.

2. As stated in claim 1, this invention provides an automated segmentation technique for deep learning for the purpose of detecting and delineating COVID-19 infections in CT images.

3. As stated in 1 and 2, this suggested innovation will need less time and money to identify the infection and annotate the infection regions as compared to existing techniques.

4. We assert that the developed method identifies infection by segmenting the lung organ from the CT image as a potential source of infection and then segmenting the infections contained inside it.

5. As stated in 1, 2, and 4, this innovation will assist physicians in determining the course of the covid-19 illness, as well as quantifying the disease's burden and severity.