DENG ENTROPY AND INFORMATION DIMENSION FOR COVID-19 AND COMMON PNEUMONIA CLASSIFICATION

Abstract

Motivated by previous authors’ work, where Shannon entropy, box covering and information dimension were applied to quantify pulmonary lesions, this paper extends such a contribution in two fashions: (i) Following the approach to quantify pulmonary lesions with Deng entropy and Deng information dimension obtained through box covering method; (ii) exploiting the Shannon and Deng lesion quantification for pulmonary illnesses classification with a bidirectional Long Short Term Memory (bLSTM). The referred pulmonary illnesses are Common Pneumonia (CP) and COVID-19. Shannon entropy and information dimension are performed here and called the Shannon sequence. Then, Deng entropy and Deng information dimension are computed for chest Computed Tomography (CT) images to obtain and combine two data sequences to quantify the pulmonary lesions. The data sequence resulting from the data combination is called the Deng sequence. Both Shannon and Deng sequences are independently used as input for the bLSTM. CT lung scans of 531 healthy subjects, 497 confirmed COVID-19 diagnoses and 516 with CP were analyzed to obtain the Shannon and Deng sequences. The results demonstrate that Deng entropy and Deng information dimension of CT images can differentiate similar lung lesions between COVID-19 and CP. Besides, a statistical analysis shows that: (a) Classification by the bLSTM is better when using the Deng sequence than the Shannon sequence; (b) Deng sequences plus bLSTM significantly outperform DenseNet-201, GoogLeNet and MobileNet-v2 in classifying COVID-19, CP and Normal CT (healthy subjects) in time and accuracy. Hence, the Deng sequence and bLSTM are fast and accurate tools for helping in diagnosing CP and COVID-19.