Automated image classification of chest X-rays of COVID-19 using deep transfer learning

Sara Dilshad; Nikhil Singh; M Atif; Atif Hanif; Nafeesah Yaqub; W A Farooq; Hijaz Ahmad; Yu-Ming Chu; Muhammad Tamoor Masood

doi:10.1016/j.rinp.2021.104529

Automated image classification of chest X-rays of COVID-19 using deep transfer learning

Results Phys. 2021 Sep:28:104529. doi: 10.1016/j.rinp.2021.104529. Epub 2021 Jul 28.

Authors

Sara Dilshad¹, Nikhil Singh¹, M Atif², Atif Hanif³, Nafeesah Yaqub², W A Farooq², Hijaz Ahmad⁴, Yu-Ming Chu^{5

6}, Muhammad Tamoor Masood⁷

Affiliations

¹ Department of Statistics and Operations Research, Aligarh Muslim University, Aligarh, Uttar Pradesh, India.
² Department of Physics and Astronomy, College of Science, King Saud University, Riyadh 11451, Saudi Arabia.
³ Botany and Microbiology Department, College of Science, King Saud University, Riyadh 11451, Saudi Arabia.
⁴ Section of Mathematics, International Telematic University Uninettuno, Corso Vittorio Emanuele II, 39, Roma 00186, Italy.
⁵ Department of Mathematics, Huzhou University, Huzhou 313000, China.
⁶ Hunan Provincial Key Laboratory of Mathematical Modeling and Analysis in Engineering Changsha, University of Science & Technology, Changsha 410114, China.
⁷ Institute for Materials Research and Innovation, University of Bolton, Bolton, United Kingdom.

Abstract

Introduction: In December 2019, the city of Wuhan, located in the Hubei province of China became the epicentre of an outbreak of a pandemic called COVID-19 by the World Health Organisation. The detection of this virus by rRTPCR (Real-Time Reverse Transcription-Polymerase Chain Reaction) tests reported high false negative rate. The manifestations of CXR (Chest X-Ray) images contained salient features of the virus. The objective of this paper is to establish the application of an early automated screening model that uses low computational power coupled with raw radiology images to assist the physicians and radiologists in the early detection and isolation of potential positive COVID-19 patients, to stop the rapid spread of the virus in vulnerable countries with limited hospital capacities and low doctor to patient ratio in order to prevent the escalating death rates.

Materials and methods: Our database consists of 447 and 447 CXR images of COVID-19 and Nofindings respectively, a total of 894 CXR images. They were then divided into 4 parts namely training, validation, testing and local/Aligarh dataset. The 4th (local/Aligarh) folder of the dataset was created to retest the diagnostics efficacy of our model on a developing nation such as India (Images from J.N.M.C., Aligarh, Uttar Pradesh, India). We used an Artificial Intelligence technique called CNN (Convolutional Neural Network). The architecture based on CNN used was MobileNet. MobileNet makes it faster than the ordinary convolutional model, while substantially decreasing the computational cost.

Results: The experimental results of our model show an accuracy of 96.33%. The F1-score is 93% and 96% for the 1st testing and 2nd testing (local/Aligarh) datasets (Tables 3.3 and 3.4). The false negative (FN) value, for the validation dataset is 6 (Fig. 3.6), for the testing dataset is 0 (Fig. 3.7) and that for the local/Aligarh dataset is 2 . The recall/sensitivity of the classifier is 93% and 96% for the 1st testing and 2nd testing (local/Aligarh) datasets (Tables 3.3 and 3.4). The recall/sensitivity for the detection of specifically COVID-19 (+) for the testing dataset is 88% and for the locally acquired dataset from India is 100%. The False Negative Rate (FNR) is 12% for the testing dataset and 0% for the locally acquired dataset (local/Aligarh). The execution time for the model to predict the input images and classify them is less than 0.1 s.

Discussion and conclusion: The false negative rate is much lower than the standard rRT-PCR tests and even 0% on the locally acquired dataset. This suggests that the established model with end-to-end structure and deep learning technique can be employed to assist radiologists in validating their initial screenings of Chest X-Ray images of COVID-19 in developed and developing nations. Further research is needed to test the model to make it more robust, employ it on multiclass classification and also try sensitise it to identify new strains of COVID-19. This model might help cultivate tele-radiology.

Keywords: Automated diagnostics; COVID-19 (coronavirus); Chest X-Ray (CXR); Deep learning; Image classification; Radiology images; Transfer learning; rRT-PCR.