Encrypted login | home

Program Information

Stability of Radiomics Feature--A Volume Dependency Study

no image available
R Luo

R Luo, J Wang , J Gan ,G Qing, H Zhong , P Hu , L Shen , W Hu ,Y Sun , Z Zhang* , Fudan University Shanghai Cancer Center, Shanghai,China, Shanghai Medical College, Fudan University, Shanghai, China


WE-G-201-9 (Wednesday, August 2, 2017) 4:30 PM - 6:00 PM Room: 201

Purpose: To investigate the stability of radiomics features regarding ROI volume.

Methods: Three image series were collected for the radiomics stability test regarding ROI volume, including a CT series of a home-made phantom, a CT series of a patient with normal liver and a MR series of a rabbit with tumor. The phantom consists of 3 cartridges filled with columnar activated carbon in different size. For each material, single slice ROIs with different size were contoured by automatically shrinkage. The image x-resolution and y-resolution for phantom, patient and rabbit are 1.067mm*1.067mm, 0.705mm*0.705mm and 0.390mm*0.390mm respectively. 40 radiomics features were extracted from these ROIs using our home-made software QIAT. For each radiomics feature, a curve of feature value versus ROI volume was plotted. Local poly-fitting was implemented to gain slope and Pearson Correlation Coefficient was calculated for each ROI volume in MATLAB. The first ROI volume with local slope≤threshold and Pearson correlation coefficient≥0.9 was considered stable volume threshold for radiomics features with certain imaging resolution.

Results: Most radiomics features are unstable in small ROI. For phantom, most features become stable when ROI reaches 3 cm in length of side, which contains more than 800 pixels. For patient normal liver, stable ROI reaches 3 cm in diameter, which contains more than 1400 pixels. For rabbit tumor, stable ROI reaches 1 cm in diameter, which contains at least 500 pixels. Some GLRLM radiomics features, for example RLN, GLN and RP, are linearly dependent on ROI volume.

Conclusion: A stable ROI for radiomics study should include enough pixels to depict whole picture of texture. Some radiomics features are volume-dependent, which might be redundant in radiomics study. Test to define volume threshold for certain texture with certain imaging resolution is recommended for stable and convincing radiomics results.

Contact Email: