ESTRO 2024 - Abstract Book

S2727

Interdisciplinary - Global health

ESTRO 2024

Ninety-seven radiation oncologists from 23 radiotherapy centers in 22 Low-and-Middle-Income Countries (LMICs) were enrolled and divided in two groups receiving different combinations of teaching and AI-assistance in contouring. This abstract reports the results from the 44 participants in the control group, who first received teaching and subsequently got access to AI-assisted contouring, as described below. The participants contoured eight organs-at-risk for four head-neck cancer patients. The cases were performed sequentially with two weeks to complete each. A mandatory virtual teaching session was held in-between the first two cases, in which the participants contoured manually. The third case was contoured with AI-assistance. Finally, the fourth case was also contoured with AI-assistance after a 6 months follow-up period (figure 1, Study Design). The teaching session was lead by an ESTRO School contouring specialist. It consisted of a two-hour discussion of the contouring guidelines used in the study (guideline [1] for brain stem and [2] for mandible, oral cavity, right submandibular gland, right parotid gland, right optic nerve, thyroid and spinal cord ). Only CT was made available to the participants, and all contouring took place in EduCase™ (RadOnc eLearning Center, Inc, Jackson, WY, USA). The AI-contours were generated using Contour+, Guideline Based Segmentation Solution, MVision AI Oy, Helsinki, Finland. The Contour+ model was based on the same guidelines as participants were instructed to use. Contour quality was quantified using the volumetric Dice Similarity Coefficient (DSC) between participants’ contours and expert consensus contours. Expert consensus contours were generated in the following steps: 1) three head-neck specialists independently contoured all structures for all cases. 2) The three sets of contours were merged using the Simultaneous Truth and Performance Level Estimation (STAPLE) [3]. 3) STAPLE-maps were binarized (threshold: 0.8). 4) The resulting contours were reviewed and edited for artifacts by an external head neck ESTRO School contouring specialist in consensus with the teaching session instructor (figure 1). The effect of teaching was assessed by comparing contour quality at baseline and after teaching. The effect of using AI-assisted contouring was assessed by comparing immediate and long-term follow-up to contouring done manually after the teaching session. Medians are reported along with 95% confidence intervals estimated with bootstrapping (two-sided, bias corrected and accelerated, 9999 iterations). Statistical difference between groups was estimated with the Mann Whitney-U test. Statistical difference of variances was determined with Levene's test. P-values below 0.05 were considered significant and are denoted *. P-values below 0.01 are denoted **.