ESTRO 2024 - Abstract Book

S5042

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Material/Methods:

Methods: A total of 97 recurrent NPC patients were enrolled, including their pre-treatment multi-sequence MR images (T1, T1C and T2 sequences) and planned re-irradiation therapy dose distribution. Multi-dimensional combinatorial features were extracted from preprocessed MRI images. Chi-square test and lasso regression were used to select the MRI-radiomics features to obtain sparse feature data. 3DCNN deep learning network model was used to combinate the selected MRI-radiomics features and dose-omics features. Eight necrosis prediction models based on T1 sequence, T2 sequence, T1C sequence, T1+T1C+T2 multi-sequence, T1+Dose, T1C+Dose, T2+Dose and T1+T1C+T2+Dose were established for training. The training data set and test data set were 82 cases and 19 cases respectively. ACC, AUC and F1-Score are used to evaluate the performance and prediction accuracy of each deep learning network model.

Results:

Results: There were 32 MRI features related to recurrent nasopharyngeal carcinoma necrosis. The 3DCNN deep learning model based on T1C single sequence and (T1+T1C+T2) multiple sequences could better predict necrosis, with the indexes of ACC, AUC and F1-Score of 0.7 and 0.7, 0.69 and 0.66, 0.63 and 0.63, respectively. The model combined MRI and dose features was more accurate for the prediction of nasopharyngeal, especially the model of multi sequence MRI plus dose features (T1+T1C+T2+Dose) shown the best performance, with the ACC, AUC and F1-Score of 0.85, 0.80 and 0.81, respectively.