Author: Yuli Lu, Chendong Ni, Cheng Qian, Kun Qian, Weiwei Sang, Chunhao Wang, Fan Xia, Zhenyu Yang, Fang-Fang Yin, Rihui Zhang, Haiming Zhu 👨🔬
Affiliation: Jiahui International Hospital, Radiation Oncology, Duke University, Medical Physics Graduate Program, Duke Kunshan University, Duke Kunshan University, The First People's Hospital of Kunshan 🌍
Purpose: To develop a radiomic quantification framework to evaluate the effects of radiomic image preprocessing hyperparameters (i.e., image resampling and discretization) on texture characterization performance.
Methods: The study employed 251 CT scans of a Credence Cartridge phantom (consisting of 10 texture materials) with different image acquisition parameters. Each material was segmented using a predefined cylindrical mask. The pre-processing workflow included 4 weighting norms(Manhattan, Euclidean, Infinity, and no weighting), 10 resampling methods(no resampling, linear, nearest neighbor, and cubic interpolation at 1mm³, 3mm³, and 5mm³), and 16 discretization methods(fixed bin sizes and counts as powers of 2, ranging from 2 to 256). Total 75 radiomic texture features(24 GLCM-based, 16 GLRLM-based, 16 GLSZM-based, 14 GLDM-based, and 5 NGTDM-based) were extracted to characterize the textural attributes. A 10-class classification experiment for materials was conducted using machine-learning models (XGB, RF and SVM). Additionally, phantoms with different intensity ranges(small~200, medium~700, and large~1200) were created using materials. Each intensity range involved four different phantoms. The same workflows were applied to each range, and four-class classification tasks were conducted. The models’ performance was evaluated using macro-AUC with a 5-fold cross-validation.
Results: Three models(XGB, RF, and SVM) successfully classified the 10 materials, achieving macro-AUC scores of 0.9924±0.0140, 0.9917±0.0141, and 0.9918±0.0148, respectively. An increasing performance trend was noted as the original CT was discretized over a larger gray level range, with performance improvements of 0.0312 as bin sizes decreased from 256 to 2 and 0.0138 as bin counts increased from 2 to 256. Among 10 resampling methods, resampling CT to 1mm³ isotropic voxel spacing using cubic interpolation showed the best performance(0.9985±0.0009). No statistically significant differences were found across the five folds. Similar trends were observed in the other three experiments.
Conclusion: The proposed framework effectively quantified the influence of hyperparameter configurations of image preprocessing on radiomics texture characterization.