Session Item

When treating cervical cancer patients with VMAT, due to sensitivity to inter-fraction motion, adaptive strategies could be required to maintain target coverage. Manual segmentation of CTV and OARs on repeat-CT scans (reCTs) is time consuming and may delay decisions whether to perform plan adaptations. Our aim was to investigate differences in clinical decisions for re-planning based on manual vs automated dose accumulation.

Twelve cervical cancer patients were included in a prospective study undergoing 5 weekly reCTs. The primary and para-aortic lymph node target and OAR volumes were manually segmented on all CTs. Clinical re-planning was performed for 3 patients. Two-arcs VMAT plans for 25 fractions of 1.8 Gy were made on the planning CT scan (planCT) and recalculated as well as robustly evaluated on the reCTs [1]. Deformable hybrid intensity and structure based image registrations were performed using the manually segmented GTV, vagina, uterus and lymph nodes CTV as controlling ROIs (DIR_manual) and without controlling ROIs (DIR_automated). The target and OAR contours were warped using DIR_automated and resulting automated versus manually segmented dose and volume differences were evaluated. The voxelwise minimum (vox min) reCT doses were warped to the planCT using DIR_manual and DIR_automated, and the different accumulated doses were compared, where the criterion for acceptable coverage was ITV D98 > 95%. 

Manual vs automated segmented vox min lymph nodes D98 (Gy) at reCTs showed a significant difference (Table 1). For the other target volumes no significant dose differences were found. Regarding OAR doses, only bowel bag Dmean (Gy) showed a significant difference. Manual vs automated segments showed only a good concordance for lymph nodes, bone marrow and sacrum (concordance index ≥ 0.8). For 7 patients, conclusions on accumulated vox min D98 ITV target coverage were identical using DIR_automated vs DIR_manual. For 2 patients, DIR_manual showed correctly that ITV coverage was not maintained in contrast to using DIR_automated (Figure 1A, marked circular). For 1 patient’s evaluation, ITV coverage was correctly maintained using DIR_manual in contrast to using DIR_automated. Out of 10 clinically re-planned reCTs, manual in contrast to automated segmentation showed correctly that individual coverage was not maintained for 4 GTV and vagina, 3 uterus and 2 lymph nodes target volumes (Figure 1B, marked circular). The ratio between automated vs manual segmented OAR volumes (cc) was close to 1 for all OARs, except for bladder and rectum (Figure 1C). Automated contour warping of the bowel bag resulted in underestimation of the DVH (Figure 1D). 

Replacing manual with automated CTV and OAR delineation could lead to missing the required re-planning and inadequate estimation of individual target and OAR doses. Deep learning CT segmentation and registration could be promising to improve the quality of dose accumulation in a time effective manner.