2304 - Radiomic Feature Stability Across Normalization Methods and Magnetic Resonance Imaging (MRI) Modalities

Purpose/Objective(s): Radiomics is an emerging field that reveals the predictive features of spatial heterogeneity of cancers. Although radiomics holds great promise in personalized medicine, the robustness of the radiomic features is critical for the predictive modeling based on it to be reproducible and generalizable. Magnetic resonance imaging (MRI) is an imaging modality that is crucial for cancer diagnosis and treatment response assessment. However, the signal variations inherent in MRI generation pose challenges for applying Radiomics to MRI data. Research so far has studied stable radiomic features for specific cancers. However, few studies have examined stable radiomic features that are shared across normalization methods and different MRI modalities, which may help to accurately describe and predict tumors on imaging.

Materials/Methods: A total of 93 T2-weighted MRI images with prostate cancer, 75 T1-weighted MRI images with pancreatic cancer, and 126 contrast-enhanced T1-weighted images with brain cancer were examined during our study using PyRadiomics. We applied the Least-squares fit tissue mean and Nyul normalization method to each MRI scan and extracted the same 924 features from the original MRIs and normalized MRIs. Stability was noted for the radiomic features if they had an intraclass correlation coefficient of 0.75 or higher.

Results: For T2-weighted MRI images of prostate cancer, 161 radiomic features were found to be stable. When examining T1-weighted MRI images of pancreatic cancer, 282 radiomic features were stable. Looking at contrast T1-weighted images with brain cancer yielded 90 radiomic features that were stable. Across the three MRI modalities, 78 stable features were identified with Intraclass Correlation Coefficients of 0.75 or higher. Stable functions and their classes are noted within Table 1 shown below.

Conclusion: From our study, 161, 282, and 90 stable radiomic features were found for prostate cancer, pancreatic cancer, and brain cancer respectively using the Nyul and Least-squares fit tissue mean normalization methods. 78 stable radiomic features were also shared across the three MRI modalities. This supports previous findings that stable radiomic features exist within cancer types. Our findings highlight stable radiomic features found across normalization methods and MRI modalities. Overall, we contribute towards the progress of forming radiomic models with MRI and other imaging modalities to predict cancer progression and response to treatment.