ESTRO 2023 - Abstract Book

S1902

Digital Posters

ESTRO 2023

Figure 1 . The study flowchart.

Results Model FL outperforms the locally trained models (Model sep ) and performs comparably to a model trained on the combined datasets (Model com ). In addition, we simulated a daily clinical scenario and embedded new module based on physician usage feedback, which demonstrated the update capability of the platform we built.

Dataset-1 (95%CI)

Dataset-2 (95%CI) 0.60 (0.47-0.70) 0.64 (0.55-0.74) 0.73 (0.63-0.85) 0.62 (0.49-0.72) 0.72 (0.63-0.82) 0.76 (0.67-0.87)

Dataset-3 (95%CI) 0.68 (0.57-0.80) 0.68 (0.55-0.79) 0.72 (0.6-0.83) 0.59 (0.48-0.71) 0.71 (0.58-0.82)

Model

AUC Model sep 0.64 (0.51-0.74)

AUC Model com 0.67 (0.57-0.72)

0.72 (0.64-0.82) 0.72 (0.64-0.81) 0.66 (0.57-0.72) 0.68 (0.6-0.76)

AUC Model FL

ACC Model sep

ACC Model com

ACC Model FL 0.81 (0.69-0.92) Table 1 . The discrimination performance of models.