Dosimetric Impact of Artificial Intelligence (AI)-Based Autocontouring Software, OncoStudio, in High-Risk Prostate Cancer Treatment Planning: A Three-Group Comparative Study on the Slice Ranges of Seminal Vesicles

TL;DR

This study compared AI-generated and manual contouring for prostate cancer treatment planning and found no significant dosimetric differences between the two methods.

Contribution

The study evaluates the dosimetric impact of AI-based autocontouring for seminal vesicles in prostate cancer treatment planning.

Findings

No statistically significant differences were observed in dose coverage between AI autocontoured and manually contoured plans.

Effect size assessments showed moderate effects in AI-based contouring, but differences were less than 1% of the prescription dose.

Results suggest AI autocontouring may be comparable to manual contouring, though larger studies are needed for confirmation.

Abstract

Objective: This pilot study investigated the dosimetric impact of artificial intelligence (AI)-generated seminal vesicle (SV) autocontouring (AC) compared with manual contouring (MC) in high-risk prostate cancer volumetric-modulated arc therapy planning.

Methods: We retrospectively analyzed treatment plans for 15 patients with high-risk prostate cancer who received 76 Gy in 38 fractions in intensity-modulated radiation therapy. Three plans were created for each patient: MC, AC with slice-adjustment to a clinical standard (2 cm from the prostate base), and unadjusted AC. Subsequently, three groups were made for each contouring method: the adjusted AC group, the unadjusted AC group, and the MC group, each including 15 plans. Primary endpoints were dose coverage of the planning target volume (PTV) and clinical target volume (CTV), evaluated by Dmax, Dmin, Dmean, D95, D98, D99, and V95%. Statistical analysis was performed using Dunnett's test, and multiple comparisons were made by selecting the MC as the control group.

Results: CTV Dmean showed a tendency to be lower in the unadjusted AC group compared with the MC group, but did not reach statistical significance (difference: -58.7 cGy, 95% confidence interval (CI): -123.0 to 5.5 cGy, p = 0.077, Hedges' g = 0.78). PTV Dmean also showed a decreasing tendency in the unadjusted AC group, but similarly did not reach statistical significance (difference: -92.8 cGy, 95% CI: -193.6 to 8.0 cGy, p = 0.075, g = 0.67). In the adjusted AC group, there was a slight tendency for a decrease in CTV Dmean, but this did not reach statistical significance (difference: -41.3 cGy, p = 0.254, g = 0.63). Effect size assessment revealed a moderate effect in both the unadjusted and the adjusted AC groups. No statistically significant differences were observed among the three groups in any dose volume histogram parameters.

Conclusion: This preliminary study (n = 15) suggests no statistically significant dosimetric differences between AI autocontoured and manually contoured SVs, with dose differences <1% of the prescription. However, given the small sample size, validation in larger cohorts is needed before clinical implementation.

Figures