ESTRO 2024 - Abstract Book

S5016

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

models were constructed respectively. In the models based on the first research endpoint (TPS < 1% vs. TPS ≥ 1%), the joint model, which includes ECOG score and skewness, performed the best, with an Area Under the Curve (AUC) of 0.722 and an Akaike Information Criterion (AIC) of 132.74, which indicated that the joint effects of clinical features and imaging radiomics features can more comprehensively reflect PD-L1 expression in NSCLC patients. Furthermore, the joint model had the lowest AIC, which meant that it had the least over-fitting, thus making it advantageous in terms of promotion and versatility. In the models based on the second research endpoint (TPS < 10% vs. TPS ≥ 10%), the Imaging radiomics model [including Volume density (ellipsoid), Cluster Prominence, and Long run high gray level emphasis] performed best, with an AUC of 0.743 and an AIC of 108.24. The calibration curve suggested that both models have good calibration degrees. In models built based on the second research endpoints, the joint model was consistent with the Imaging radiomics model, which meant that the inclusion of clinical features did not improve the performance of the model, possibly because the information contained in images more comprehensively reflects PD-L1 expression in NSCLC patients, which also indicates that PD-L1 expression can be predicted by pre-treatment enhanced CT images of NSCLC patients alone.