Document Type : Original Paper
Authors
1
Medical Physics Department, School of Medicine, Iran University of Medical Sciences, Tehran, Iran
2
Radiation Science Department, Faculty of Allied Medicine, Iran University of Medical Sciences, Tehran, Iran
3
Department of Radiation Physics, Unit 1420, The University of Texas MD Anderson Cancer Center, 1515 Holcombe Boulevard, Houston, TX 77030, USA
4
Medical Physics Department, School of Medicine, Iran university of Medical Sciences,
10.22038/ijmp.2026.93264.2657
Abstract
Purpose: The reproducibility of radiomic features could be a serious obstacle that limits further applications. This study aims to assess the reproducibility of MRI radiomic features by using a biological ovarian phantom and different acquisition parameters.
Materials and Methods: Three 1.5 Tesla MRI scanners from different manufacturers as the first source, and also alterations in imaging parameters, including slice thickness, space between slices, image weight, and fat saturation sequence, as the second source of feature variations, were evaluated. In addition, to evaluate the effect of image normalization on feature reproducibility, all the images were normalized. Ninety-three radiomic features from 6 feature classes, including First-Order, GLCM, GLDM, GLRLM, GLSZM, and NGTDM, were calculated by the 3D-Slicer. Reproducibility of features was measured by COV, ICC, and CCC.
Results: The significant impact of scanner and image weight variation on feature reproducibility is obvious when about 90% and 64% of features showed 20 %< COV, respectively. On the other hand, slice thickness was the least affected source, where 58.8% of features showed excellent reproducibility (COV ≤ 5%). GLRLM showed the best reproducibility against scanner variation (ICC=0.6996 and CCC=0.3503). Also image normalization has positively affected feature reproducibility in the scanner variation scheme. Additionally, good (5 %< COV≤10%) and intermediate (10 %< COV≤20%) COV groups have increased by normalization.
Conclusion: MRI radiomic features are highly dependent on image acquisition scanner types and imaging parameters, and utilizing biological phantoms can lead to reliable outcomes that make the way of clinical translation of these results easier. Future works should be the priority in the robustness evaluation of radiomic features, and the inconsistent behavior of the image normalization filter needs higher attention.
Keywords
Main Subjects
)