Session Item

The aim of this study was to evaluate the selection of the plan-of-the-day (PoD) in CBCT-guided adaptive radiation therapy (ART) of locally advanced cervical cancer.

Data from 49 patients with cervical cancer included in a phase II ART clinical trial (ARCOL) was collected from 7 clinical centers. For each patient, three planning CT scans (158 CT scans in total) were acquired with different bladder fillings (empty, intermediate and full) resulting to three plans available to generate a treatment plans library. A CBCT was acquired every day (1217 CBCTs in total) for a radiation oncologist to select visually the best plan and the patients were treated.  

The PoD selection was evaluated retrospectively by computing the geometrical coverage of the daily CTV (on the CBCT scan) by the CTV of the planning CT scans. This coverage was defined as the percentage of the daily CTV included in the planning CTV. For this purpose, a deep learning model (nnU-Net) was trained, on a separate database, to segment the main structures in CBCT scans. Each resulting automatically segmented daily CTV was assessed by a radiation oncologist into three categories: good, intermediate and poor. From the segmented daily CTV, the treatment plan providing the higher geometrical coverage was identified (“optimal plan”) and compared to the plan actually selected during the treatment (“actual plan”). CBCT scans associated to a poor segmentation were excluded. The Wilcoxon test was used to compare the coverage of the CTV between the two approaches (“optimal” and “actual” plans), according to the quality of the segmentation.

Out of the 1217 automatically segmented CTVs, 256 CTVs were estimated as good, 674 CTVs as intermediate and 287 CTVs as poor by the physician. The PoD concordances between “optimal” and “actual” plans were 74.6% and 62.3% for the segmentations estimated as good and intermediate, respectively. The geometrical coverages of the target were significantly higher with the “optimal plans” than with the “actual plans” (median values of 85% versus 84% and 82% versus 81% for the segmentations estimated as good and intermediate, respectively (Fig 1).

Fig 1: CTV geometrical coverage of the “optimal” and “actual” treatment plans based on the quality of the CTV segmentation. ( * ∶ p_value<0.05 )

The concordance between the “actual plans” and the “optimal plans” was found to be between 62% and 74%, depending on the CTV automatic segmentation quality. However, these differences resulted to a small difference in terms of geometrical coverage of the target. The deep-learning segmentation method could be used to help the physicians to select the PoD during treatments.