A framework for quantitative analysis of Computed Tomography images of viral pneumonitis: radiomic features in COVID and non-COVID patients

TL;DR

This study developed and validated an AI-based pipeline using radiomic features and quantitative metrics from CT images to differentiate COVID-19 from other viral pneumonias with high accuracy.

Contribution

The paper introduces a comprehensive pipeline combining radiomic analysis and machine learning for viral pneumonia classification, optimized during the COVID-19 pandemic.

Findings

Radiomic features achieved up to 81% accuracy in classification.

Gaussian model accurately described healthy lung tissue in 94% of cases.

AI models demonstrated good diagnostic performance across the dataset.

Abstract

Purpose: to optimize a pipeline of clinical data gathering and CT images processing implemented during the COVID-19 pandemic crisis and to develop artificial intelligence model for different of viral pneumonia. Methods: 1028 chest CT image of patients with positive swab were segmented automatically for lung extraction. A Gaussian model developed in Python language was applied to calculate quantitative metrics (QM) describing well-aerated and ill portions of the lungs from the histogram distribution of lung CT numbers in both lungs of each image and in four geometrical subdivision. Furthermore, radiomic features (RF) of first and second order were extracted from bilateral lungs using PyRadiomic tools. QM and RF were used to develop 4 different Multi-Layer Perceptron (MLP) classifier to discriminate images of patients with COVID (n=646) and non-COVID (n=382) viral pneumonia. Results: The Gaussian model applied to lung CT histogram correctly described healthy parenchyma 94% of the patients. The resulting accuracy of the models for COVID diagnosis were in the range 0.76-0.87, as the integral of the receiver operating curve. The best diagnostic performances were associated to the model based on RF of first and second order, with 21 relevant features after LASSO regression and an accuracy of 0.81$\pm$0.02 after 4-fold cross validation Conclusions: Despite these results were obtained with CT images from a single center, a platform for extracting useful quantitative metrics from CT images was developed and optimized. Four artificial intelligence-based models for classifying patients with COVID and non-COVID viral pneumonia were developed and compared showing overall good diagnostic performances