Investigating the use of artificial intelligence for daily CBCT‐guided online adaptive radiotherapy

External beam radiotherapy is an effective and organ-preserving treatment of cancer, used for many disease sites. To account for anatomical changes occurring during the course of radiotherapy, safety margins are added to the target area in order to ensure dose coverage of the disease. These margins, however, also lead to irradiation of healthy tissue. Because of large day-to-day variations in bladder, rectum, and bowel volume and position, large margins are conventionally needed for treatments in the pelvis. With the use of online adaptive radiotherapy (oART), where the treatment is adapted to the anatomy of the day, the margins can be reduced, thereby sparing healthy tissue, while maintaining target coverage. Until recently, time-consuming procedures have limited the clinical implementation of this approach, but through technological advancements, including increased computer power and integration and use of artificial intelligence (AI), previous time-consuming procedures have been automated and speeded up. A novel system using AI to perform cone beam computed tomography (CBCT) guided oART was in September 2019 implemented at Herlev Hospital. Consequently, expanding the role of the CBCT images to be used daily as basis for treatment planning. While this treatment technique comes with possibilities such as significant reductions in dose to healthy tissue, it also comes with challenges and uncertainties. This PhD project aimed at investigating the benefits, uncertainties, and challenges in oART through three studies:
In study I, we investigated the dosimetric benefit and clinical feasibility of oART for patients with bladder cancer. Through evaluation of clinical oART treatments, we found that patient-specific oART margins reduced the irradiated volume for all patients and significantly spared the healthy tissue (bowel and rectum). oART was deemed feasible, within a time slot of 30 min compared to 15-20 min for conventional treatments. Editing the AI-generated bladder contour was the most time-consuming task in the oART procedure and especially necessary for patients with catheter or irregular bladder position due to prostatectomy, i.e., anatomy for which the AI was not trained on. In March 2022, a clinical phase II trial was implemented to evaluate the clinical effect, primarily in terms of reduced gastrointestinal toxicity. By May 31, 2024, 19 out of 90 participants had been included in the study.
In study II, we investigated oART for patients with anal cancer. In a dosimetric evaluation of patients previously treated with a conventional non-adaptive treatment technique, we found that a proposed margin reduction in oART significantly reduced the dose to healthy tissue (bowel and bladder). Evaluation of target propagation methods revealed that editing auto-deformed targets was necessary to ensure adequate target coverage. Furthermore, rigid propagation of pelvic lymph nodes was deemed necessary to reduce the need for manual edits and due to limited visibility of the primary tumour on the CBCT. The treatment technique was implemented clinically in January 2022, with the implementation of a phase II trial to investigate the clinical effect, primarily in terms of reduced gastrointestinal toxicity. By May 31, 2024, 84 out of 205 participants had been included in the study.
In study III, we evaluated AI-generated contours and variation in target delineation among radiotherapy technologists (RTTs) performing oART of bladder cancer routinely. We found that AI-generated contours were smaller than all manually edited ones, and that it was necessary to edit AI-generated contours to obtain adequate target coverage. A margin to compensate for the manual corrections could not be suggested. The variation among RTTs was non-isotropic and similar to previous studies performed among physicians.
To conclude, we demonstrated that CBCT-guided oART is feasible and helps reduce dose to important healthy organs in patients with bladder or anal cancer. Furthermore, our findings support an RTT-driven oART workflow. While AI is essential for speeding up the process, our studies caution against unsupervised workflows using unedited AI-generated contours, as this will result in underdosage of the target for many patients. Initiated prospective trials will provide crucial clinical data.