Artificial intelligence-generated targets and inter-observer variation in online adaptive radiotherapy of bladder cancer

Lina M Åström; Patrik Sibolt; Hannah Chamberlin; Eva Serup-Hansen; Claus E Andersen; Marcel van Herk; Lene S Mouritsen; Marianne C Aznar; Claus P Behrens

doi:10.1016/j.phro.2024.100640

Artificial intelligence-generated targets and inter-observer variation in online adaptive radiotherapy of bladder cancer

Phys Imaging Radiat Oncol. 2024 Sep 1:31:100640. doi: 10.1016/j.phro.2024.100640. eCollection 2024 Jul.

Authors

Affiliations

¹ Department of Oncology, Copenhagen University Hospital - Herlev and Gentofte, Copenhagen, Denmark.
² Department of Health Technology, Technical University of Denmark, Roskilde, Denmark.
³ Division of Cancer Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom.

Abstract

Background and purpose: Daily target re-delineation in online adaptive radiotherapy (oART) introduces uncertainty. The aim of this study was to evaluate artificial intelligence (AI) generated contours and inter-observer target variation among radiotherapy technicians in cone-beam CT (CBCT) guided oART of bladder cancer.

Materials and methods: For each of 10 consecutive patients treated with oART for bladder cancer, one CBCT was randomly selected and retrospectively included. The bladder (CTV-T) was AI-segmented (CTV-T_AI). Seven radiotherapy technicians independently reviewed and edited CTV-T_AI, generating CTV-T_ADP. Contours were benchmarked against a ground truth contour (CTV-T_GT) delineated blindly from scratch. CTV-T_ADP and CTV-T_AI were compared to CTV-T_GT using volume, dice similarity coefficient, and bidirectional local distance. Dose coverage (D_99%>95 %) of CTV-T_GT was evaluated for treatment plans optimized for CTV-T_AI and CTV-T_ADP with clinical margins. Inter-observer variation among CTV-T_ADP was assessed using coefficient of variation and generalized conformity index.

Results: CTV-T_GT ranged from 48.7 cm³ to 211.6 cm³. The median [range] volume difference was 4.5 [-17.8, 42.4] cm³ for CTV-T_ADP and -15.5 [-54.2, 4.3] cm³ for CTV-T_AI, compared to CTV-T_GT. Corresponding dice similarity coefficients were 0.87 [0.71, 0.95] and 0.84 [0.64, 0.95]. CTV-T_GT was adequately covered in 68/70 plans optimized on CTV-T_ADP and in 6/10 plans optimized on CTV-T_AI with clinical margins. The median [range] coefficient of variation was 0.08 [0.05, 0.11] and generalized conformity index was 0.78 [0.71, 0.88] among CTV-T_ADP.

Conclusions: Target re-delineation in CBCT-guided oART of bladder cancer demonstrated non-isotropic inter-observer variation. Manual adjustment of AI-generated contours was necessary to cover ground truth targets.