Real world AI-driven segmentation: Efficiency gains and workflow challenges in radiotherapy.

BACKGROUND

It remains unclear whether improving contouring efficiency using AI-driven contouring (AIseg) significantly shortens the OAR contouring task time in a real world setting, or the overall radiotherapy planning-CT to treatment time. This single institution multidisciplinary study aims to evaluate how AIseg changes the duration taken for contouring tasks as well as the time to complete the overall treatment planning carepath from planning CT to treatment start.

METHODS

This retrospective study evaluated both palliative and radical radiotherapy OAR contouring time metrics in a large, real-world patient cohort across four years. Data included both conventional and ablative radiation schedules. Data on task availability, initiation, and completion were recorded from ARIA patient records across a four-year period: three years before (pre-AI) and one year after AI implementation (post-AI). "In-progress" OAR contouring times (from task initiation to completion) and OAR contouring workflow times (from task availability to completion) were analysed across multiple anatomical sites, including head and neck, thorax, abdomen, breast, and pelvis. Trends were assessed monthly to determine if any immediate (step) or gradual (slope) changes associated with AIseg introduction occurred. Additionally, overall CT-to-treatment intervals were evaluated to see if contouring efficiencies translated into shorter CT-to-treatment workflows.

RESULTS

A total of 9,964 pre-AI and 3,820 post-AI OAR contouring "in-progress" tasks were analysed, alongside 16,352 pre-AI and 5,870 post-AI "workflow" tasks. AIseg consistently reduced median active contouring times 51.5 % (p < 0.001), and up to 70 % in the most complex cohorts (e.g., head and neck, thorax). Month-by-month trend analyses showed that prior to AIseg, contouring workflow times trended upward. Post-AIseg, these same trends gradually improved and sloped downward (p < 0.001). Despite these notable gains at the task and workflow levels, there was no corresponding decrease in overall CT-to-treatment intervals. However, in the post-AI period, a significantly higher proportion of plans were approved and ready for treatment.

CONCLUSION

AIseg offers substantial efficiency gains in active contouring, particularly for complex cases, and resulted in increasing improvements in contouring workflow times over the post-implementation period. Although overall CT-to-treatment timelines remained unchanged due to fixed scheduling constraints, a significantly greater proportion of plans were ready earlier post-AIseg implementation. Our study challenges the assumption that task-level efficiencies automatically translate into faster overall patient treatment pathways, underscoring the critical need for deliberate workflow and scheduling optimisation to ensure that time savings yield meaningful improvements in patient timelines and outcomes.