Deep learning empowered COVID-19 diagnosis using chest CT scan images for collaborative edge-cloud computing platform

Abstract

The novel coronavirus outbreak has spread worldwide, causing respiratory infections in humans, leading to a huge global pandemic COVID-19. According to World Health Organization, the only way to curb this spread is by increasing the testing and isolating the infected. Meanwhile, the clinical testing currently being followed is not easily accessible and requires much time to give the results. In this scenario, remote diagnostic systems could become a handy solution. Some existing studies leverage the deep learning approach to provide an effective alternative to clinical diagnostic techniques. However, it is difficult to use such complex networks in resource constraint environments. To address this problem, we developed a fine-tuned deep learning model inspired by the architecture of the MobileNet V2 model. Moreover, the developed model is further optimized in terms of its size and complexity to make it compatible with mobile and edge devices. The results of extensive experimentation performed on a real-world dataset consisting of 2482 chest Computerized Tomography scan images strongly suggest the superiority of the developed fine-tuned deep learning model in terms of high accuracy and faster diagnosis time. The proposed model achieved a classification accuracy of 96.40%, with approximately ten times shorter response time than prevailing deep learning models. Further, McNemar's statistical test results also prove the efficacy of the proposed model.

Keywords: COVID-19; Chest CT scan; Deep Learning; Diagnosis; Edge Computing; MobileNet V2.

Fig. 1

Representation of newly registered cases of COVID-19 infection across India from April 2020 to August 2020

Fig. 2

Radar plot comparing the edge and cloud computing technologies in terms of privacy, resources, latency, storage, and reliability metrics

Fig. 3

Typical connected smart healthcare system architecture, which includes IoT devices, edge layer, and cloud layer

Fig. 4

Knowledge transfer process in transfer learning

Fig. 5

The proposed transfer learning framework for screening of COVID-19 infection using chest CT scan images

Fig. 6

Concept of Depthwise separable convolutions

Fig. 7

Basic building block consisting of an expansion layer, a depthwise convolution, and a projection layer in MobileNet V2 architecture

Fig. 8

The proposed MobileNet V2 based fine-tuned model. The figure depicts the transfer of knowledge learnt from the ImageNet dataset for the application of COVID-19 diagnosis

Fig. 9

Collaborative edge cloud computing framework

Fig. 10

Few sample images of chest CT scan from the dataset

Fig. 11

Representation of confusion matrices of each transfer learning model for the prediction of COVID-19 infection on the test data

Fig. 12

Comparison of overall performance in terms of accuracy, precision, F-1 score, and MCC values across VGG 16, VGG 19, DenseNet 201, and proposed model

Fig. 13

Comparison of sensitivity and specificity scores of each transfer learning

Fig. 14

Representation of performance in terms of ROC curve for all developed transfer learning models. The plot is zoomed from the top left corner for better visualization

Fig. 15

Analysis of the time complexity for each transfer learning model in terms of average time taken for classification of test data images

Fig. 16

Analysis of the hardware complexity for each transfer learning model in terms of disk size taken by the trained model

Fig. 17

Few sample inferences obtained from the model using Python Interpreter

Fig. 18

Representation of contingency matrix developed in McNemars statistical test