Volume change of the prostate during moderately hypo-fractionated radiotherapy assessed by artificial intelligence.

Background: Diagnostic quality MRI acquired daily for radiotherapy (RT) planning on an MR-linac allows longitudinal evaluation of the patients' anatomy. This study investigated changes in prostate volume during MR-guided RT. The changes were assessed from manual delineations used clinically for daily online adaptation as well as automated segmentation by artificial intelligence (AI). The consistency and congruity of these two methods were evaluated. Methods: The prostate volumes were extracted from daily planning MRI scans of 45 patients receiving 60 Gy in 20 fractions. These volumes were manually edited during the online adaptive treatment planning workflow. The prostate was re-segmented retrospectively for each fraction by AI with an in-house developed nnU-net, trained on prostate cancer patients. The volume for each fraction was normalized to the volume at the patients' 1st fraction to identify possible time trends. Results: Increased population mean prostate volume was seen both based on manual and automatic segmentation. However, based on manual delineations, the peak volume occurred at the 12th fraction at 106.8% of the initial volume, while based on automatic segmentation, the volume peaked at a mean increase 110.8% by the 5th fraction. Standard deviation of volumes for automated segmentation (5.2%) versus manual delineation (12.7%), and reduced variation between fractions from 3.6% to 2.6% indicate better consistency of the automatic segmentation. Conclusion: Automated segmentation by our locally trained nnU-net was more consistent than manual delineations performed clinically. The population mean increase in prostate volume peaked at 110.8% by the 5th fraction after reduce over the remaining treatment course.