Ultrasonic-Based Radiomics Signature With Machine Learning for Differentiating Prognostic Subsets of Pediatric Peripheral Neuroblastic Tumors: A Retrospective Study

Objective

To construct and select a better model based on ultrasonic-based radiomics features and clinical characteristics for prognostic subsets of pediatric neuroblastic tumors.

Methods

Data from 73 children with neuroblastic tumors were included and divided into a training group and a validation group. Data 1 contained the subjects’ radiomics features and clinical characteristics, while data 2 contained radiomics features. With the help of machine learning, five models were constructed for data 1 and data 2, respectively. The model with the highest accuracy and area under the curve was selected as the combined model and radiomics model for data 1 and data 2, respectively. A superior model was then chosen from the models after further comparison.

Results

The extreme gradient-boosting model for data 1 was chosen as the combined model and the extreme gradient-boosting model for data 2 was chosen as the radiomics model. The area under the curve of the combined and radiomics models in the validation group was 0.941 and 0.918 (p = 0.6906). The balanced accuracy, kappa value and F1 score of the radiomics model (0.9045, 0.8091 and 0.9091, respectively) were higher than those of the combined model (0.8545, 0.7123 and 0.8696, respectively). The top eight features of the radiomics model included five first-order statistical features and three textural features, all of which were high-dimensional features.

Conclusion

Our study proved that the radiomics model outperformed the combined model at differentiating prognostic subsets of pediatric neuroblastic tumors. Additionally, we found that high-dimensional ultrasonic-based radiomics features surpassed other features and clinical characteristics.