Development and Clinical Evaluation of Quantitative Lung Fibrosis Scores on CT Images in Clinical Trials: Idiopathic Pulmonary Fibrosis and Scleroderma Lung Disease

• Activity Number: 219 - Artificial Intelligence and Machine Learning in Medical Imaging for Drug Development
• Type: Invited
• Date/Time: Wednesday, August 11, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Section on Statistics in Imaging
• Author(s): Grace Hyun Kim* and Jonathan Goldin and Matthew Brown
• Companies: UCLA and UCLA and UCLA
• Keywords: Artificial Intelligence ; Machine Learning; IPF; Lung Fibrosis ; Clinical trials; COVID-19

Abstract

Model development, analytic evaluation, and clinical evaluation are important steps to demonstrate the robustness of model from AI. Model development and Evaluation: Radiomic features derived from the intensities of CT have been shown to be discriminative, yet the features are not reproducible in different CT parameters and models. To normalize the variation, the denoised features were extracted from mathematically decomposed CT images by the total variation. We developed a quantitative lung fibrosis (QLF) via machine learning: (a) denoise, (b) feature selection by smoothly clipped absolute deviation, and (c) classifier by support vector machine. We evaluated the scores at 25%, a meaningful threshold of survival and found good performance (AUC=0.96). Clinical evaluation: moderate to good correlations were found between QLF and the clinical outcomes of forced vital capacity at baseline and the changes (rho>0.40) in 10 clinical trials. Changes in QLF have been used to support efficacy or predict the progression in subjects with Idiopathic pulmonary fibrosis (IPF) and scleroderma. Moreover, this technique has been extended for infectious diseases such as Tuberculosis and COVID-19.