Characterization of Radiomics Features Extracted from Images Generated by the 0.35 T Scanner of an Integrated MRI-Linac

Sammanfattning: Purpose: In an era of personalized oncology where the aim is to give every patient the right treatment at the right time an area of promising research is emerging called radiomics, or quantitative image analysis. The main underlying hypothesis is that pathophysiological information can be found in image texture not visible to the bare eye that can improve diagnosis, treatment adaption or be linked to a certain clinical outcome. This project in- vestigates the stability and repeatability of radiomic features extracted from images acquired with an integrated MRI-Linac with a 0.35 T scanner. The main objective was to identify radiomic features that are robust over various imaging conditions in both phantom and human data. Methods: The patient dataset included 50 images from ten stereotactic body radiation therapy (SBRT) pancreas cancer patients treated with 5 fractions, given on a daily basis, on the integrated MRI-Linac. Two anatomical sites were selected to represent heteroge- neous invariant tissue: the kidneys and liver. Eleven images from a Magphan RT phantom and 11 images from a ViewRay Daily QA phantom acquired monthly and daily respec- tively, constituted the basis for the phantom data, representing ideal imaging conditions. All images were acquired with a True Fast Imaging with Steady State Free Precession (TRUFI) pulse sequence with two different protocols. A high resolution (1.5mm3 voxel resolution) protocol was used for all phantom images and a protocol with lower resolution (1.5mm2 x 3.0mm) was used to collect the patient images. Totally 1087 shaped-based, first order statistics, second order statistics and higher order statistical radiomic features were extracted from each region of interest (ROI) and subject. Stability was assessed with the Coefficient of Variation (CoV) where features with CoV<5% were classified as robust. Common robust features among all datasets were identified as a final step. Results: There were in total 130 radiomic features demonstrating robustness (CoV<5%) among all datasets. Robust features could be identified within each category, apart from two second order statistics groups: Gray level size zone and Neighborhood gray tone difference. The mean value of the CoV and the corresponding standard deviation was calculated for each robust feature in all four datasets. Discussion and Conclusion: Several robust features in common with the result of this work can be identified in other MRI-based radiomics studies, which is promising. However, no overall agreement is found between all studies, emphasizing the need of more stability assessment research. The result in this work indicates that robust radiomic features over various imaging conditions, in both phantom and patient data, can be identified. It im- plies that phantom measurements can be used in stability assessment studies and that the 0.35 T scanner of the integrated MRI-Linac in this work is sufficiently stable over timefor radiomic studies. An additional promising finding is that many robust features also have been reported to have predictive value or discriminative power in other studies. Al- though preliminary, this result can serve as guidelines for further model building or further