Online adaptive radiotherapy (oART) requires organs at risk (OAR) and target contours to be created on a daily basis. The accuracy of these contours is pivotal for successful and timely oART. This study evaluates the quality of artificial intelligence (AI) auto-contouring in cervical cancer using the Varian Ethos Therapy system. 

Five historic patients with cervical cancer were chosen from a previous departmental study.  All had received 50.4Gy in 28 fractions with daily soft tissue matching using CBCT on a Varian Truebeam and had OAR (bladder, rectum, bowel, uterus) contoured on each daily CBCT by a radiation oncologist. Using the full bladder planning CT scan, a new 12-field IMRT plan was created using the Varian Ethos TPS, and 5 weekly CBCTs for each patient were uploaded onto a software emulator. OART was simulated twice, once with the AI-generated OAR contours edited by a clinician and once with the contours left unedited. Time taken to edit the contours in the emulator was recorded.  The contours were exported to the Eclipse planning system, where they were compared to the original contours quantitatively via dice similarity coefficient (DSC) and 95-percentile Hausdorff distance (HD95%). The unedited contours were rated qualitatively from excellent through to unacceptable. 

The DSC and HD95% had statistically significant differences between the AI and manually edited contours for all apart from the rectum structure. Box plots of the results for each are given in Figure 1.

The bladder had the highest DSC for the AI and manually edited contours with a mean of 0.93 and 0.94 respectively (see Table 1). The Bladder and Rectum had the smallest HD95% for the AI at 4.32mm and 5.43mm respectively. The bowel was the least accurate performing AI generated structure using the DSC and HD95% metrics. Manually editing the bowel still only yielded a mean DSC of 0.79 whereas all other structure achieved over 0.85. The manually edited contours were considered acceptable, differences between these and the original contours are due to inter-observer variability; hence this also accounts for part of the difference between the AI-generated contours and the originals.

Qualitative assessment of the AI generated contours showed that the bladder and rectum were rated excellent or good in 96% of cases, requiring minimal editing. The bowel contouring was rated unacceptable in 16% of cases. The average time taken to edit the AI-generated contours was 12:20 minutes (06:20-17:35 minutes), with the bowel taking on average 07:30 minutes.

Initial experience has shown that auto-generated contours for the female pelvis are satisfactory, and that the time required for manual editing is feasible for most cases within an oART workflow.  Bowel was the least acceptable structure and required considerable human adjustment. Further evaluation is underway to investigate the use of a bowel bag contour to reduce time required for bowel contour editing.