ESTRO 2024 - Abstract Book

S4178

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

[5] K. I. Kauweloa et al., “GateCTTM surface tracking system for respiratory signal reconstruction in 4DCT imaging,” Med Phys, vol. 39, no. 1, pp. 492–502, 2012, doi: 10.1118/1.3671941.

659

Poster Discussion

Daily adaptive workflow for bladder cancer radiotherapy: can AI do the complete job online?

Sana Azzarouali, Karin Goudschaal, Jorrit Visser, Laurien Daniëls, Arjan Bel, Duncan den Boer

Amsterdam UMC, Radiation oncology, Amsterdam, Netherlands

Purpose/Objective:

Manual recontouring during online adaptive radiotherapy (oART) can be labor intensive and time consuming, potentially leading to additional intrafractional organ variability that might need to be compensated for by a larger margin. A fully automated workflow without the need for human intervention would tackle these problems and allow for shorter session times thus more patient comfort and less intrafractional bladder filling. Additionally, more consistency can be expected and the treatment quality would be less dependent on staff experience. Our aim was to evaluate the geometric and dosimetric consequences of an artificial intelligence (AI) based fully automated oART workflow for bladder cancer without human interference.

Material/Methods:

Seventeen patients with muscle-invasive bladder cancer were treated with a CBCT-guided oART workflow in 20 fractions. A total dose of 40 Gy was given to the bladder, urethra and pelvic lymph nodes (CTV elective ), and the tumor bed (CTV boost ) received an additional simultaneous integrated boost (SIB) dose of 15 Gy (VMAT). A GTV-CTV boost and CTV boost -PTV margin of 5 mm were used. The CTV elective -PTV margin was at least 7 mm [1]. AI was used to delineate the bladder and rectum, so-called influencers, on the daily CBCT. These influencers were manually corrected when deemed necessary and were used for a structure-guided deformable registration to propagate the target and organ at-risk delineations. After correcting/accepting these delineations, a treatment plan reoptimization was performed leading to the adaptive plan. To study the efficacy of the manual corrections to the delineations, all clinical adaptive plans (plan clin ) used during online treatment, for which manual corrections to the structure delineations were allowed, were compared with simulated adaptive plans from fully automated sessions (plan auto ). Plan auto was obtained by simulating the oART workflow on the same CBCTs as used online for plan clin but without manual corrections to any of the delineations. The geometric differences were evaluated by extracting the Dice Score (DSC), 95% Hausdorff distance (95%HD) and the relative volume of the bladder, rectum and CTV boost delineations from plan auto (V auto ) and plan clin (V clin ). The dosimetric consequences were assessed by propagating the clinical structures to plan auto and extracting the target coverage, i.e. volume receiving 95% of the prescribed dose (V95%). The target coverage was also extracted for the clinical structures on plan clin and a statistical comparison was done by performing a Wilcoxon signed rank test. Furthermore, we assessed the frequency of sessions that met the clinical requirement for the target coverage (V95%≥98%) for both CTV boost and CTV elective .