CT-Based Radiomics and Neural Network Approaches for Predicting Wilms Tumor Recurrence

Background/Objectives : Wilms tumor is the most common kidney childhood tumor. Predictive markers for better understanding the tumor behavior are merged. Radiomics as noninvasive quantitative prognostic biomarkers may help to predict tumor recurrence. The study aimed to specify a Wilms tumor recurrence radiomics signature. Methods : Patients diagnosed with Wilms tumor and treated from 2011 to 2024 in Western Ukrainian Specialized Children's Medical Center in Lviv, Ukraine, were included in the study. Demographic and clinical characteristics of patients were obtained from the patient’s records. An initial abdominal CT scan was analyzed using the PyRadiomics package. Radiomics features (including wavelets) of the primary tumor were analyzed using a neural network (NN) algorithm. SPSS (v27.) built-in and Tensor Flow manually built neural networks architecture were applied. ROC curves analysis was applied to estimate the model performance. Results : Sixty-three Wilms tumor patients with a median age of 32.1 months (IQR 17.4-57) were included in the exploratory analysis. Median follow-up time was 45 months (IQR 21.5-84.5). Recurrence was registered in 9 patients (14.3%). For each case, 810 Radiomics features were analyzed and provided to the NN model, which was trained on 52 patients and tested on 12 patients. The Tensor Flow manually built NN architecture model reached an accuracy of 77% for prediction of tumor recurrence. The area under the ROC curve (AUC) for SPSS analysis was 0.789. Conclusions : CT-based radiomics signature shows an acceptable performance in the prediction of Wilms tumor recurrence. A similar performance was demonstrated in two different NN architectures. Further external validation of the model is needed.