ESTRO 2024 - Abstract Book

S4526

Physics - Machine learning models and clinical applications

ESTRO 2024

Figure 1: Proposed Standardisation Framework

Results:

Table 1 shows the weighted average F1-scores for internal and external validation centres. It was observed that the F1 scores of the multimodal approach were consistently better than the univariate models for the internal and external validation centres, providing evidence for the effectiveness of the VisualGPT-based multimodal feature fusion approach in modelling the discriminative characteristics of the RT structures. The use of spatial relationships also contributed to capturing the context of an RT structure including its surrounding and relative position information instead of using only shape/geometric information. Variations in the relative origin points (0,0,0) of the scan between patients, similar shape features of the PTVs and CTVs may have led to inaccuracies in the geometric features.

Table 1: Weighted average F1-scores for internal and external validation centres