ESTRO 2024 - Abstract Book

S3044

Physics - Autosegmentation

ESTRO 2024

Figure 1 provides a visual comparison of the performance of the different models.

Model A performed inconsistently, with performance drops for three of the main chambers for Patient 1 and generally poor performance for Patient 2 (LVAD), c.f. Figure 2. Model B improved scores with respect to model A, but still failed to create accurate contours for Patient 2 and struggled with two chambers for Patient 1. For Patient 3, models A and B perform on par with the inter-observer variability for the four chambers and pulmonary artery but not for other structures. Model B showed increased performance for training data-providing centres compared to those who did not. Models C/C* significantly increased the auto-contouring performance for all test patients. Here, the performance for the four chambers, major vessels and valves was non-inferior compared to manual inter-observer variability (p<0.05). Table 1 provides a summary of the performance for all structures.