COVID-19 infection analysis framework using novel boosted CNNs and radiological images

COVID-19, a novel pathogen that emerged in late 2019, has the potential to cause pneumonia with unique variants upon infection. Hence, the development of efficient diagnostic systems is crucial in accurately identifying infected patients and effectively mitigating the spread of the disease. However, the system poses several challenges because of the limited availability of labeled data, distortion, and complexity in image representation, as well as variations in contrast and texture. Therefore, a novel two-phase analysis framework has been developed to scrutinize the subtle irregularities associated with COVID-19 contamination. A new Convolutional Neural Network-based STM-BRNet is developed, which integrates the Split-Transform-Merge (STM) block and Feature map enrichment (FME) techniques in the first phase. The STM block captures boundary and regional-specific features essential for detecting COVID-19 infectious CT slices. Additionally, by incorporating the FME and Transfer Learning (TL) concept into the STM blocks, multiple enhanced channels are generated to effectively capture minute variations in illumination and texture specific to COVID-19-infected images. Additionally, residual multipath learning is used to improve the learning capacity of STM-BRNet and progressively increase the feature representation by boosting at a high level through TL. In the second phase of the analysis, the COVID-19 CT scans are processed using the newly developed SA-CB-BRSeg segmentation CNN to accurately delineate infection in the images. The SA-CB-BRSeg method utilizes a unique approach that combines smooth and heterogeneous processes in both the encoder and decoder. These operations are structured to effectively capture COVID-19 patterns, including region-homogenous, texture variation, and border. By incorporating these techniques, the SA-CB-BRSeg method demonstrates its ability to accurately analyze and segment COVID-19 related data. Furthermore, the SA-CB-BRSeg model incorporates the novel concept of CB in the decoder, where additional channels are combined using TL to enhance the learning of low contrast regions. The developed STM-BRNet and SA-CB-BRSeg models achieve impressive results, with an accuracy of 98.01%, recall of 98.12%, F-score of 98.11%, Dice Similarity of 96.396%, and IOU of 98.85%. The proposed framework will alleviate the workload and enhance the radiologist's decision-making capacity in identifying the infected region of COVID-19 and evaluating the severity stages of the disease.


Introduction
The new coronavirus (COVID-19) is a transmissible disease that first appeared in December 2019 and spread worldwide 1 . COVID-19 is an ongoing pandemic that has devastatingly affected the world 2 . The COVID-19 suspected cases are approximately 675 million, with 6.8 million deaths, while 647 million have been healthier. It is estimated that 99.6% of the infected patients have slight, while 0.4% have severe or critical symptoms 3 . However, it causes respiratory inflammation, difficulty breathing, pneumonia, alveolar damage, and respiratory failure in severe cases, eventually leading to death 4 . The person with COVID-19 pneumonia mostly depicts the signs of pleural effusion, ground-glass opacities, and consolidation 5 .
The detection tests for COVID-19 include molecular testing (RT-PCR) and Chest radiological imaging (X-ray, CT scan) [6][7][8] . Chest imaging is also used to complement clinical evaluation, monitoring, and follow-up for COVID-19 diagnosed patients. Moreover, the CT scan is utilized for the severity assessment and treatment of COVID-19 patients. In a public health emergency, the manual examination of many radiological images is a great challenge and a severe concern for remote areas without experienced radiologists 9 . Radiological images usually are complex.
The COVID-19 infected region has high variation in size, shape, and position. Furthermore, these radiological images are highly distorted due to noise during CT image acquisition 10 Automatic detection technology is a serious need to help radiologists improve their performance and deal with many patients and will overcome the burden of manually examining. Therefore, Deep Learning (DL) based diagnostics techniques are developed to facilitate radiologists in identifying COVID-19 infection 11 . Such an effective predictive model can overcome the radiologist burden for manual assessment of COVID-19 infected CT, ultimately improving the survival rate. The contribution of DL and its capability to classify and segment the image with high accuracy will eliminate the probability of incorrect results by the currently used testing kits.
DL will reduce the load on healthcare facilities 12,13 .
The DL-based automated technique's remarkable success in different fields has attracted researchers to its application in medical diagnostic systems 14 . These tools are designed for automatic medical image analysis and facilitate radiologists in identifying lung-related anomalies 15 . These tools can detect minor irregularities of COVID-19 patterns that cannot be observed through a manual examination and reduce the burden on hospitals for COVID-19 diagnosis.

COVID-19 Infection Detection
The proposed detection phase constitutes two modules (i) the proposed STM-BRNet detection  the reception field and preserve data dimensions at the output layer to achieve a diverse feature set to distinguish infected regions from healthy areas 25 . Moreover, the new CB concept is altered 7 at the STM blocks to preserve the reduced prominent maps and then joined to get various boosted channels and capture minor infection contrast variation. In addition, using various pooling operations results in down-sampling that eventually enhances the robustness of the model against any variation. Additionally, the region operator within the STM block utilizes the average pooling layer for smoothening and noise reduction.

Architectural Design of the Proposed STM-BRNET
The STM-BRNet comprises two STM blocks with identical topology and is arranged methodically to learn various features' initial and final levels. Each STM-RENet comprises four convolutional blocks, where Region and boundary operations are methodically employed. The dimension of each STM boosted block is 256 and 512 26,27 . As the prime focus of the architecture is to get minor contrast and texture infection patterns, therefore, four diverse blocks, namely Region and Edge (RE), Edge and Region (ER), Edge (E), and Region (R), are implemented. The dilated convolutional layer, regional/boundary operations, and CB idea are altered to learn COVID-19 specific features in each block.
The RE block extracts regions and boundaries; it comprises two dilated convolutional layers followed by the average and max-pooling layers, as shown in equations (1-3). Moreover, the ER block extracts edges and regions; it comprises two dilated convolutional layers followed by a max-pooling layer. The E and R block learn the edges and smoothness, respectively. In block E, additional channels are generated by TL to achieve various channels, while block RE, ER, and E are learning from scratch.
These processes enhance the boundary information and region-specific properties, whereas dilated convolutional operations aid in learning the global receptive features. The perception of multipath-based STM blocks is used to obtain diversity in the feature set and can dynamically capture the minor representative and textural variations information from the COVID-19 infected CT images. Moroever, fully connected layers and dropout layers that preserve the prominent features and reduce overfitting.
The channels and size are represented by x and k x l. The kernels and their size are denoted by f and i x j in equation (1). In contrast, the output ranges to [1 to k-m+1, l-n+1]. Moreover, average and max-pooling window size is represented by w, respectively, on convolved output (x , ) (Equations 2-3). In equation (4), the original feature maps of block RE, ER, and R are signified by x RE , x ER , and x R , respectively. Likewise, the auxiliary feature-maps of block R achieved using TL are denoted as x E . These channels are boosted by concatenation operation b(.). The neuron quantity and activation in equation (6) are shown with v a . and σ.

Implementation of existing Detection CNNs
Recently, CNN has demonstrated effective performance in medical field images to detect and segment medical images 14 . The employed models for detection are VGG-16/19, ResNet-50, ShuffleNet, Xception, etc. 28 . These deep CNNs with varying in-depth and network designs are tailored to detect and segment COVID-19 infected radiological images. Moreover, the initial and final layers are customized according to the target specific domain.

COVID-19 Infected Regions Segmentation
The proposed STM-BRNet aims to classify COVID-infected patients from healthy patients by utilizing the capabilities of deep CNN architectural ideas. The infected images are provided the segmentation CNNs for delineating COVID-19 infection regions that identify the disease's severity. This paper implements two different experimental setups for infection segmentation: (i) proposed SA-CB-RESeg segmentation, (ii) target-specific segmentation CNNs implementation from scratch, and TL.

Proposed SA-CB-RESeg Segmentation CNN
An SA-CB-RESeg is proposed to perform fine-grain pixel-wise segmentation. The proposed SA-CB-RESeg CNN is comprised of two encoders and boosted decoder blocks. The encoder and decoder blocks are designed in such a way as to improve the SA-CB-RESeg learning capacity. In this regard, average-pooling and max-pooling, along with convolutional operation in encoding and decoding stages, are employed systematically to learn region and boundary-related properties of COVID-19 infected regions 29,30 . Moreover, the convolutional operation employed a trained filter on images and generated feature maps of distinctive patterns. The encoders and decoders are designed symmetrically; however, in pooling operations, max-pooling is employed in the encoder for down-sampling. Contrarily, in the decoder, an un-pooling operation is employed to perform up-sampling. Finally, a 2x2 convolutional layer is employed to classify pixels into COVID-19 and background.
The encoder is designed to learn semantically meaningful COVID-19 specific patterns.
However, the encoder loses spatial information essential for infected region segmentation because it reconstructs the infection map. In this regard, decoders are employed to preserve the spatial information of the corresponding encoders using pooling indices. These positional indices are stored in each pooling operation and are helpful for reconstruction and mapping on the decoder side. Moreover, the pooling operation performs down-sampling and reduces the spatial dimension ( Figure 3).

Static Attention
Static attention (SA) enhances the learning capability of the COVID-19-infected areas by locating high weightage. The detail of the SA block is illustrated in Figure 4. demonstrates the input channel and is the pixel-weightage coefficient at the range of [0, 1] (Equation (10)).
The output _ highlights the infected region while suppressing the irrelevant features. In Equations (11) and (12), 1 and 2 is the Relu and Sigmoid activation function, respectively.

Implementation of Existing Segmentation CNNs
Several deep CNNs are employed to segment the COVID-19 CT infected region using diverse datasets 31 . This study employs implemented DeepLab, U-SegNet, SegNet, VGG-16, U-Net, and FCN as segmentation models [32][33][34] . The existing segmentation CNNs have been implemented for comparative studies. We have employed the existing CNN models by training from scratch and weight initialization. The weights are initialized from pre-trained CNNs using the concept of TL and fine-tuned on CT images.

Dataset
Chest CT scan has a high sensitivity for the diagnosis of COVID. The major benefit of using lung CT scans is that it makes the internal anatomy more apparent as overlapping structures are

Implementation details
The detection and segmentation CNNs are trained separately in the developed diagnosis system. selecting optimal hyper-parameters for smooth and efficient convergence 36 . The hyperparameters detail is available in Table 1.

Performance evaluation
The developed framework's performance is evaluated using standard measures, and its detail is illustrated in Table 2. The detection measures include accuracy, recall, etc., depicted in Equations (13)(14)(15)(16)(17). While the segmentation CNNs are assessed using IoU and DS coefficient that is expressed in Equations (18) and (19), respectively. Segmentation accuracy (S_Acc) to correctly predictction of positive and negative class samples. In comparison, S-Acc is used for the correct prediction of pixels. DS metric is used for structure similarity, and IoU is employed to identify the predicted vs. ground truth's overlapping ratio.

Mathew Correlation Coefficient
MCC Identify the quality of confusion metrics on an unbalanced dataset.

Jaccard Coefficient
IoU %The similarity between Label and predicted areas.

Dice-Similarity
DS % The weighted_similarity between label and predicted areas.

Segmentation-Acc
S_Acc %Pixels that are accurately partitioned into COVID-19 and Background.

Results
This paper proposes a new two-stage diagnosis framework to analyze the COVID-19 infectious region in the lungs. Distributing the proposed into two stages has two main advantages: improving the performance and reducing the computational complexities. Moreover, screening of COVID-19 infected samples and then analyzing the infectious region helps quickly identify the severity of the disease. Furthermore, the two-stage process rivals the clinical workflow, where patients are referred for further diagnostic tests after initial detection. The performance of the proposed STM-BRNet detection and SA-CB-RESeg segmentation CNNs are evaluated based on standard performance metrics. The proposed models are tested on unseen data and indicate considerable performance compared to existing CNNs.

In this stage, a deep CNN-based STM-BRNet is developed to detect COVID-19-infected images.
This stage is optimized with a high detection rate for recognizing the COVID-19 characteristic pattern and reducing false positives (Table 3). The learning ability of STM-BRNet for COVID-19 specific CT images is assessed and compared with existing CNNs. We optimized this stage for a high detection rate for recognizing the COVID-19 characteristic pattern with fewer false positives (shown in Table 3

The Proposed STM-BRNet's Performance Analysis
The proposed STM-BRNet is assessed on the test set using several performance measures like  (Table 3). The performance of the STM-BRNet is further increased by adding fully connected and dropout layers to emphasize the learning and improve the generalization.

Performance analysis with the existing CNNs
The proposed STM-BRNet performance is compared with the five customized classification CNNs (VGG-16/19, ResNet-50, Xception, and ShuffleNet). The customized CNNs are famous for solving complex challenges and may be successively used to identify lung abnormalities. For a fair comparison, customized CNNs have been learned about the COVID-19 specific image. In contrast, the proposed STM-BRNet shows outperformance and performance gain in F-score, MCC, accuracy, etc., with the customized CNNs on the test dataset, as shown in Table 3 and Figure 9.

Features Visualization and PR/ROC Analysis
The considerable detection ability of STM-BRNet is evident from the principal components analysis (PCA) plot. PCA can be used to reduce the dimensionality of STM-BRNet features and 17 identify the distinctive patterns for better discrimination. For comparison, deep feature-based analysis of best performing existing ResNet-50 is also provided in Figure 7. The considerable learning ability of the developed STM-BRNet is evident from the PCA plot that includes the first, second, and third generated principal components. Moreover, detection rate curves (PR/ROC) are also used to quantitatively assess the discrimination ability of the developed STM-BRNet ( Figure 8). These are performance measurement curves that evaluate the generalization of the STM-BRNet by analyzing the discrimination between two COVID-19 infected and healthy classes at different threshold setups. Moreover, STM-BRNet has a good learning ability compared with different CNNs on the optimal threshold. The PR and ROC curve for COVID-19 detection based on deep STM-BRNet features gives a higher AUC, indicating better model performance.

Infected Region Analysis
CTs infected are separated using the developed STM-BRNet and assigned to deep segmentation CNN to analyze the infectious region. The infected slices and normal have minor contrast variations in the early stage. However, isolating the infected region from the healthy region is quite challenging. Therefore, the proposed SA-CB-RESeg segregates the infected regions by identifying infection boundaries and has minor contrast variation. Moreover, region analysis is needed to identify the severity of the mild, medium, or severe disease and its treatment design.

Segmentation analysis of the proposed SA-CB-RESeg
The SA-CB-BRSeg is developed to segment COVID-19 infectious regions in CT lung images.
The existing segmentation CNNs have been optimized based on COVID-19 infected specific patterns and imagery features. The experimental results on unseen data show the significance of the proposed SA-CB-BRSeg (Table 4) (Table 4). In comparison, they precisely learned the discriminative boundaries and achieved a higher value of BFsa (99.09 %).

Segmentation Analysis with the Existing CNNs
The existing segmentation CNNs are employed to evaluate the learning capacity of the proposed SA-CB-BRSeg. In this regard, SA-CB-BRSeg performance is compared with six popular segmentation CNNs (DeepLabv3, U-SegNet, SegNet, U-Net, VGG-16, and FCN) ( Table 4 and The proposed SA-CB-BRSeg appears globally suited for moderate to severely infected regions. Moreover, the proposed and existing models' performance is improved using radiological and augmented data. The developed SA-CB-RESeg has low complexity and in-depth but shows more accurate performance than highly complex and large-depth models. Incorporating pixel-wise distribution of the developed SA-CB-RESeg improved the segmentation for various stages of infected regions. The performance metrics in Table 4 and the visual quality of the segmented maps in (Figure 10) evidence the outperformance of the proposed SA-CB-RESeg.  37 Infected 91.00 20 Infected 19 Infected Gl-Acc, Mn-Acc. represents global and mean accuracy where Mn-IoU and Wt-IoU denote means and weighted IoU.

FIGURE 9
The proposed STM-BRNet and SA-CB-BRSeg performance gain over the existing CNNs.  38,39 . Moreover, the radiologist labeled and augmented data are combined to improve the proposed SA-CB-RESeg performance 40 . The segmentation for the analysis of infected regions is achieved using the best-performing existing TL-based trained DeepLabv3

Visual Analysis of the Proposed SA-CB-BRSeg
Visual analysis of COVID-19 infection segmentation using deep SA-CB-BRSeg is used to identify and analyze infected regions. The subjective evaluation shows that the proposed SA-CB-BRSeg accurately highlights the infected region. Incorporating pixel-wise distribution of the proposed SA-CB-BRSeg improved the segmentation of various stages of infected areas. The performance metrics in Table 5   CT. The proposed SA-CB-RESeg benefited from training from scratch and fine-tuning on COVID-19 data using TL and CB. The integrated approach discovers the whole COVID-19

Conclusions
infected region, which may help the radiologist estimate the disease's mild, medium, and severe stages. In contrast, a single-phase framework may not effectively give a precise and accurate detailed analysis of the infected region. COVID-19 is a novel infectious disease, and publically available labeled samples are limited. Therefore, in the future, we will employ the proposed framework on big datasets to improve the reliability of real-time diagnostics. Moreover, the dataset can be increased using the augmentation of the training sets by generating synthetic examples using GAN. Furthermore, it may be modified to segregate the infectious region into multi-class characteristic patterns automatically.

Conflicts of interest
Authors declared no conflict of interest.

Availability of data and material
Publicly available dataset is used in this work that is accessible at https://medicalsegmentation.com/covid19/ https://zenodo.org/record/3757476#.YVyHNdpBw2w