Work

Using a physics-based approach to standardize radiomics values: experimental validation in an anthropomorphic phantom on a clinical CT scanner using a range of dose levels and reconstruction kernels

Year: 2026

Type: conference-paper

Abstract: Radiomics relies on quantitative features to discern the underlying biological signatures. However, feature dependence on the imaging systems themselves hampers the creation of reproducible and generalizable models. We have previously proposed a novel framework to remove the effects of system blur and image noise on radiomic calculations and performed validation in simulation studies. In this work, we extended the analysis and evaluated the method on CT data acquired of an anthropomorphic phantom with realistic lung textures. Data was acquired at five different dose levels and reconstructed using eight different reconstruction kernels. To test the generalizability of the method, we applied our proposed method to standardize from all possible starting and reference kernel pairs under all measured dose levels for a total of 320 cases (8×8×5). Standardization was performed for radiomics features from four classes, histogram, GLCM, gray-level run length matrix-(GLRLM), and wavelet transforms. Results indicate that standardized radiomics features are closer to the reference and on average, the average absolute percentage difference from reference over all features is improved by a factor of three compared with unstandardized features. In addition, we found that standardization from a smoother to a sharper kernel is a more challenging task and that performance is comparable across all dose levels. This work shows that the proposed standardization method is effective in standardizing radiomics feature values across a wide range of imaging conditions in clinical CT.

Source: SPIE Medical Imaging: Physics of Medical Imaging

DOI: 10.1117/12.3087679

Authors: Huay Din, Yijie Yuan, Grace Hyun Kim, Michael McNitt-Gray, J. Webster Stayman, Grace J. Gang

Institutions: University of Pennsylvania, Johns Hopkins University, University of California, Los Angeles

Language: English

Topic: Radiomics and Machine Learning in Medical Imaging

Subfield: Radiology, Nuclear Medicine and Imaging

Field: Medicine

Domain: Health Sciences

SDG: Good health and well-being

Open Access status: green

Funder: National Institutes of Health