ESTRO 2024 - Abstract Book

S2246

Clinical - Upper GI

ESTRO 2024

Material/Methods:

We analyzed 112 EC patients aged ≥65 years who received definitive RT between 2010 and 2018. Pre -treatment image data (CT and PET images), treatment planning information (dose distribution, segmentation data), clinical information and patient outcome data were collected. The gross tumor volume (GTV) of primary lesion, GTV of primary lesion plus lymph node metastases, the clinical target volume, and the planned target volume were delineated. These regions of interest (ROIs) were further expanded and shrank and a total of 23 analysis regions were created as the ROIs of tumor regions. Furthermore, we created four skeletal muscle regions at the level of the third lumbar vertebra and segmented 13 lung function regions based on CT values. A total number of 67,520 radiomics features per a patient image were extracted from CT and PET images. We performed dimension reduction using the least absolute shrinkage and selection operator (LASSO) analysis to select the relevant features for prognosis prediction. A Rad-score was generated using Cox regression analysis based on these selected features. Five prediction models were built: four radiomics model (tumor region, skeletal muscle region, skeletal muscle/lung regions and tumor/skeletal muscle/lung regions) and clinical model. Nomograms were created for each model, and Kaplan-Meier analysis was conducted. We evaluated the model accuracy using the C-index. Of all cases, 70% were used for model construction, and 30% were used for accuracy validation.

Results:

In this analysis, the clinical model selected stage as a factor for the nomogram, while the radiomics model selected Rad-score. The C-index values were as follows for each model: tumor region 0.7, skeletal muscle region 0.67, skeletal muscle/lung regions 0.8, tumor/skeletal muscle/lung regions 0.86, and clinical model 0.6. Significant differences were observed in the radiomics model for the muscle/lung regions, and the radiomics model for the tumor/skeletal muscle/lung regions exhibited the highest predictive accuracy. In the survival rate curves, when the total point of the nomogram was used to classify high-risk and low-risk based on the optimal values in training, significant differences were observed between the high-risk and low-risk groups for all models. The most significant risk stratification was achieved in the radiomics model that included the tumor/skeletal muscle/lung regions (3-year overall survival rate: 83% vs. 35%).

Conclusion:

We developed a survival prediction model for elderly EC patients who underwent definitive RT. The radiomics model demonstrated higher predictive accuracy compared to the Clinical model, and the image features extracted through radiomics analysis were considered to be more significant prognostic factors than clinical factors. Furthermore, radiomics analysis using tumor regions, skeletal muscles, and lung regions improved the predictive accuracy. These results suggest the possibility of creating survival prediction models that consider survival risks beyond the primary cause of death by incorporating image features reflecting skeletal muscle quality and lung function. Our results also suggest that in elderly EC patients, in addition to cancer treatment, maintaining and improving respiratory and physical functions may also impact prognosis.

Keywords: esophageal cancer, elderly, radiomics