Improving the Reproducibility of Computed Tomography Radiomic Features Using an Enhanced Hierarchical Feature Synthesis Network

Abstract

Radiomics has gained popularity as a quantitative analysis method for medical images. However, computed tomography (CT) scans are performed using various parameters, such as X-ray dose and reconstruction kernels, which is a fundamental reason for the lack of reproducibility of radiomic features. This study evaluated whether the proposed network improves the reproducibility of radiomic features across various CT protocols and reconstruction kernels. We set five CT scan protocols and two reconstruction kernels to create various noise settings for the obtained CT images with an abdominal phantom. We developed an enhanced hierarchical feature synthesis (EHFS) network to improve the reproducibility of radiomic features across various CT protocols and reconstruction kernels. Eight hundred and nineteen radiomic features were extracted, including first-order, second-order, and wavelet features. Reproducibility was assessed using Lin’s concordance correlation coefficient (CCC) on internal and external testing. We considered a radiomic feature with CCC $\ge0.85$ as a high-agreement feature. As a result, the average number of reproducible features increased in all protocols, from 241 ± 38 to 565 ± 11 in internal testing. In external testing, consisting of a new phantom and unseen protocol, 239 ± 74 reproducible features were in source images and 324 ± 16 were in generated images. The EHFS network is a novel approach to improving the reproducibility of radiomic features. It outperforms existing methods in reproducibility and generalization, as demonstrated by comprehensive experiments on both internal and external datasets. Our deep-learning-based CT image conversion could be a solution for standardization in ongoing radiomics research.

Keywords

• Artificial intelligence
• computed tomography
• radiomics
• reproducibility