ESTRO 2021 Abstract Book

S1545

ESTRO 2021

10x5x4=200 scans we extracted 93 gray-level radiomic features for the delineated structures CTV_prostate and Femoral_head_L /R using PyRadiomics (v2.2.0). The delineations were checked by a physician and adapted when necessary on the first image set of each fraction. Only features with an average relative standard deviation (RSD) per fraction below 10% were kept to eliminate features not giving a robust signal within fractions. Linear regression on the median feature values over the five fractions measured time trends for each feature over the course of the total treatment period. For the selected structures the total number of significant trends (p<0.05) in radiomic features over all patients are reported. Results A total of 50 features had a low RSD over the fractions. The results for the linear regression test showed 140 (out of 500) significant trends for CTV_Prostate (Figure 1) , clustered in six patients. For the femoral heads only 18/16 (L/R) features had significant trends. Figure 2 shows an example feature (RobustMeanAbsoluteDeviation) for the four scans per fraction for a single patient (#4) as well as the fit to the median values per fraction. From the graphs it is clear that the CTV reveals a trend whereas the trend in Femoral_Head_L is negligible.

Conclusion During the two to three weeks MR-linac treatment, up to 20 MRI’s are routinely made of a single patient over the course of the treatment. This analysis is an indication of a time trend of radiomic feature values. The time-trends vary from patient