DOI: 10.1002/ima.23028 ISSN: 0899-9457

Differentiation of COVID‐19 pneumonia from other lung diseases using CT radiomic features and machine learning: A large multicentric cohort study

Isaac Shiri, Yazdan Salimi, Abdollah Saberi, Masoumeh Pakbin, Ghasem Hajianfar, Atlas Haddadi Avval, Amirhossein Sanaat, Azadeh Akhavanallaf, Shayan Mostafaei, Zahra Mansouri, Dariush Askari, Mohammadreza Ghasemian, Ehsan Sharifipour, Saleh Sandoughdaran, Ahmad Sohrabi, Elham Sadati, Somayeh Livani, Pooya Iranpour, Shahriar Kolahi, Bardia Khosravi, Maziar Khateri, Salar Bijari, Mohammad Reza Atashzar, Sajad P. Shayesteh, Mohammad Reza Babaei, Elnaz Jenabi, Mohammad Hasanian, Alireza Shahhamzeh, Seyed Yaser Foroghi Ghomi, Abolfazl Mozafari, Hesamaddin Shirzad‐Aski, Fatemeh Movaseghi, Rama Bozorgmehr, Neda Goharpey, Hamid Abdollahi, Parham Geramifar, Amir Reza Radmard, Hossein Arabi, Kiara Rezaei‐Kalantari, Mehrdad Oveisi, Arman Rahmim, Habib Zaidi
  • Electrical and Electronic Engineering
  • Computer Vision and Pattern Recognition
  • Software
  • Electronic, Optical and Magnetic Materials


To derive and validate an effective machine learning and radiomics‐based model to differentiate COVID‐19 pneumonia from other lung diseases using a large multi‐centric dataset. In this retrospective study, we collected 19 private and five public datasets of chest CT images, accumulating to 26 307 images (15 148 COVID‐19; 9657 other lung diseases including non‐COVID‐19 pneumonia, lung cancer, pulmonary embolism; 1502 normal cases). We tested 96 machine learning‐based models by cross‐combining four feature selectors (FSs) and eight dimensionality reduction techniques with eight classifiers. We trained and evaluated our models using three different strategies: #1, the whole dataset (15 148 COVID‐19 and 11 159 other); #2, a new dataset after excluding healthy individuals and COVID‐19 patients who did not have RT‐PCR results (12 419 COVID‐19 and 8278 other); and #3 only non‐COVID‐19 pneumonia patients and a random sample of COVID‐19 patients (3000 COVID‐19 and 2582 others) to provide balanced classes. The best models were chosen by one‐standard‐deviation rule in 10‐fold cross‐validation and evaluated on the hold out test sets for reporting. In strategy#1, Relief FS combined with random forest (RF) classifier resulted in the highest performance (accuracy = 0.96, AUC = 0.99, sensitivity = 0.98, specificity = 0.94, PPV = 0.96, and NPV = 0.96). In strategy#2, Recursive Feature Elimination (RFE) FS and RF classifier combination resulted in the highest performance (accuracy = 0.97, AUC = 0.99, sensitivity = 0.98, specificity = 0.95, PPV = 0.96, NPV = 0.98). Finally, in strategy #3, the ANOVA FS and RF classifier combination resulted in the highest performance (accuracy = 0.94, AUC =0.98, sensitivity = 0.96, specificity = 0.93, PPV = 0.93, NPV = 0.96). Lung radiomic features combined with machine learning algorithms can enable the effective diagnosis of COVID‐19 pneumonia in CT images without the use of additional tests.

More from our Archive