ESTRO 2024 - Abstract Book

S3151

Physics - Autosegmentation

ESTRO 2024

Contouring duration is crucial in adaptive workflows, often representing the main time burden in the daily multidisciplinary treatment session. In this context, automated contouring (AC) advances may significantly fasten radiation oncologists’ routine. Previous findings in terms of geometric indexes showed deep learning (DL)-based tools outperforming traditional algorithms in AC in cervical cancer radiotherapy (CC-RT). This study aimed to evaluate if the dosimetric effects of organ at risk (OAR) AC inaccuracies confirm previous evidence.

Material/Methods:

Four AC tools were evaluated: MVISION (MV, MVision AI, Helsinki, Finland) and LIMBUS AI (LI, Limbus AI Inc, Regina, Canada), Atlas-Based Auto Segmentation (STAPLE, ST, Elekta AB, Stockholm, Sweden), and Random Forest algorithm (RF, Admire, Elekta AB, Stockholm, Sweden). A mono-institutional consecutive series of 40 CC-RT patients were retrospectively selected. Twenty structure sets (SS) were randomly chosen as the atlas set for STAPLE and RF. The remaining 20 SS (testing) were auto-segmented by STAPLE, RF, MV, and LI. Original target volumes were added to the AC sets. A total of 100 SS were used to automatically optimize 100 plans (20 patients x 5 SS each) delivering 50 Gy in 25 fractions using Monaco TPS (mCycle, research v.59, Elekta AB, Sweden). To avoid any optimization and time bias, any manual intervention after automated planning optimization and calculation was forbidden. The plans obtained from the manually contoured SS (Ground Truth, GT) were used as the planning GT. Dose distributions obtained with AC SS were superimposed on the GT SS to register OAR dose differences caused by delineation variations. Plan complexity was measured in terms of monitor units (MU) and Modulation Degree (MD). Statistical significance was assessed by performing the Wilcoxon-Mann-Whitney test (p<0.05).

Results:

Automated planning failed with 2 out of 100 SS: the PTV-bowel intersection volume was too large in 1 ST and 1 MV SS leading to a TPS failure message. All but one GT automated plans resulted clinically acceptable. OAR dose differences caused by LI delineation inaccuracies did not affect plan acceptability in the whole set. On the other hand, ST, RF, and MV plans showed unacceptable target coverage in 3 out of 20 cases to preserve OAR sparing. No statistical significance was observed with small median deviations. Nevertheless, it is worth noticing that dose deviations could be significant at an individual patient level, affecting the clinical plan acceptability (e.g., PTV V9 5% -38.9% and bowel V 45Gy -90.7%). AC variations influenced plan optimization and modulation as expected: the maximum variation of median MU was registered in ST plans (+7.5 [-7.4 – 27.3] %) while a not statistically significant increase in median MD was observed in RF and ST SS, but not in DL SS.

Conclusion:

DL AC inaccuracies appeared to have less dosimetric effect than other traditional AC tools but no statistical significance has been registered. Further investigation on a larger cohort is warranted to study possible significant correlations between geometric indexes and dose deviations. CC-RT is in fact characterized by strong anatomical variability averaging out dose impact. Although AC reduces contouring time, these preliminary results proved that automated contouring validation by radiation oncologists remain crucial in the routine or in an adaptive workflow to avoid significant dosimetric effects on a single-patient basis.

Keywords: Cervical cancer, dose effects, automated planning