Patient characteristics and the clinical workflow

Between April 2021 and August 2023, 17 patients with muscle-invasive bladder cancer (see also Additional file 1) were treated with AI-driven CBCT-guided oART (Ethos Therapy™, version 1.1, Varian a Siemens Healthineers Company, USA). Within 6 weeks after TURBT the patients underwent chemotherapy (Mitomycin-c/ Capecitabine) starting on the first day of RT treatment.

RT was given over a period of 4 weeks with a total of 20 fractions. The clinical target volume (CTVelective) was defined as the pelvic lymph nodes (internal iliac, obturator, hypogastric and perivesical until lower part of sacroiliac joint), urethra (men: 2 cm proximal, women: 1 cm proximal) and whole bladder. This CTVelective received a total dose of 40 Gy (including positive lymph nodes when present). The (remnant) tumor or resection scar, which for simplicity both will be referred to as gross tumor volume (GTV), received an additional boost dose of 15 Gy given as a simultaneous integrated boost (SIB). All structures were manually delineated on a reference CT. CTVboost was defined as the GTV with an isotropic 5 mm margin. The PTV margin for the CTVboost was also 5 mm and a minimum of 7 mm was used for the PTV margin of the bladder based on the bladder filling observed during pretreatment (using 2 planning CTs). More details about the pretreatment (including placement of fiducial markers, treatment planning constraints, margins and delineation of target and OARs) were previously reported and are available online [6].

The online adaptive session started by acquiring a CBCT at the start of the daily anatomy. The AI network (vendor supplied) used this CBCT as input to automatically delineate the bladder and rectum (for more details on the AI network and automatic segmentation see [17]). In the clinical workflow, manual adjustments to these delineations were allowed (3 physicians and 9 radiation therapists were involved). Subsequently, a deformable registration was performed from the reference CT to the CBCT, used to generate a synthetic CT for dose calculation. The bladder and rectum influence the position and shape of the GTV, therefore the software used these to guide the deformable registration to come to a GTV delineation for a manual check. OARs (small bowel, bowel bag, sigmoid, left and right femur head) were propagated using deformable registration from the reference CT to the daily image. An adaptive plan was generated taking into account the daily anatomy (see also [6] for more details of the oART workflow).

Simulation of an automated oART workflow

To study the potential of an oART workflow that would be fully automated during the online sessions, two data sets were compared (see Fig. 1). The first data set consists of 340 online reoptimized treatment plans from the 17 patients (20 fractions per patient) treated in the clinic as described in the previous paragraph (Evaluationclin). Evaluationclin consists of the structure set (Contourclin) and the dose distribution (Dclin) extracted from the online fractions that included manual adjustments to the delineations if deemed necessary. The second data set was obtained by first simulating the oART workflow steps on the same daily CBCTs (Ethos test environment, version 1.1, Varian a Siemens Healthineers Company, USA). In contrast to the clinical workflow, no manual adjustments were made to the structure set, including influencers and target, automatically proposed by the software (Contourauto). A geometric evaluation was done by comparing Contourclin with Contourauto as described in the next paragraph. To also evaluate the dosimetric effects, a dose distribution was generated by performing a reoptimization based on Contourauto (Dauto). To evaluate if the simulated dose distribution would have led to acceptable treatments, the dose-volume histogram of Contourclin was calculated using Dauto (Evaluationauto), where Contourclin was used as the ground truth. For this evaluation, the clinical requirements (for details see previous work) were assessed [6].

Fig. 1

Evaluation of the treatment plan from the online fully automated daily adaptive workflow as compared to the clinically used treatment plan

Workflow comparison

We first monitored the number of manual corrections that were applied to the influencers (i.e. bladder and rectum) and GTV in the clinical workflow. To compare the online fully automated workflow with the clinical workflow, all 20 sessions from each of the 17 patients were included for evaluation. The evaluation consisted of a geometric contour comparison between Contourauto and Contourclin, a dosimetric comparison between Dauto and Dclin and an analysis of what might influence the accuracy of the online fully automated workflow. All metrics were extracted using home built software (Matlab R2021a, Mathworks).

Geometric contour comparison

To analyze the geometric differences between Contourauto and Contourclin, the Dice Similarity Coefficient (DSC), the relative volume, the 95-percentile Hausdorff Distance (95%HD) and the Mean Distance to Agreement (MDA) were extracted from each of the 340 session for the influencers and the CTVboost [18,19,20]. The relative volume was defined as Vauto/Vclin, where Vauto represents the volume of Contourauto and Vclin the volume of Contourclin.

Dosimetric and statistical analysis

A dosimetric evaluation of the online fully automated workflow was done by evaluating Contourclin of the GTV, CTVboost, CTVelective, the planning target volume (PTV) surrounding CTVboost (PTVboost) and the PTV surrounding CTVelective (PTVelective) in dose distributions Dclin and Dauto. The target coverage of these clinical contours was determined for the two dose distributions by extracting the volume of these structures receiving a minimum of 95% of the prescribed dose (V95%). The clinical requirement for the V95% was a minimum of 98% [6]. To get an indication of the healthy tissue sparing, the V95% outside the previously mentioned target structures (V95%,out) was obtained by:

where V95%,body represents the total volume of the body receiving a minimum of 95% of the prescribed dose per fraction [6]. To compare the difference between Dauto and Dclin using these metrics, a statistical analysis was done by a paired Wilcoxon signed-rank test. A Bonferroni corrected significance level of 0.5% was used. Besides performing a dosimetric comparison between Dauto and Dclin, we also tested whether the target coverage achieved with Dauto would have met the clinical requirements.

Volume differences and delineation accuracy

To get an insight of what might influence the software delineation accuracy, the variations in bladder and rectum volume were analyzed. Since the influencers guide the online GTV delineation, the volume difference between these influencers on the reference CT and the online CBCT was extracted to determine its effect on the CTVboost coverage (V95%). The number of sessions meeting the clinical requirement for the target coverage was evaluated by using volume ranges of 50 cm3 and 25 cm3 for the volume differences of the bladder and rectum, respectively. To complete the evaluation, the effect of this volume difference on the target coverage of CTVelective was also analyzed.