Prediction of early recurrence in primary intussusception: development of an ultrasound-based radiomics and deep learning nomogram

Abstract Background Intussusception recurrence frequently happens within 48 hours post-reduction, with current evaluations primarily relying on clinical symptoms and lacking efficient early prediction models. Objective This study seeks to develop a prediction model for Early Intussusception Recurrence(EIR) based on abdominal ultrasound images. Materials and methods A total of 1,314 pediatric intussusception cases, including both clinical and imaging data, were retrospectively collected from three hospitals between January 2016 and December 2024. Cases from Center 1 were randomly divided into the training set and internal validation set at a ratio of 6:4, whereas the remaining two centers were assigned as independent external test sets (Test 1 and Test 2), respectively. An integrated model combining DeepLabv3 and ResNet101 was adopted for automatic lesion segmentation, and feature extraction was conducted based on the segmented regions of interest (ROIs). Specifically, ResNet50 was employed as the backbone model in the transfer learning framework to extract deep learning (DL) features, with simultaneous extraction of radiomics features. Subsequently, the predictive performance of these combined features was systematically evaluated using three machine learning classifiers: k-nearest neighbor (KNN), random forest (RF), and extreme gradient boosting (XGBoost). Finally, a visualized nomogram was constructed for recurrence risk prediction by integrating the optimal model with key clinical variables. Results The automatic segmentation model constructed in this study exhibited stable and excellent performance in the training and test sets, with Dice coefficients of 0.935 and 0.862, and intersection over union (IOU) of 0.880 and 0.779, respectively. Statistical analysis identified age, onset-to-presentation time, and vomiting as key clinical features associated with EIR. A total of 7 radiomics features and 9 deep learning (DL) features were selected for modeling, and radiomics-deep learning fused (DLR) features were generated therefrom. Finally, a nomogram was constructed by integrating the DLR-XGBoost model with key clinical variables. This nomogram demonstrated stable and reliable predictive performance across cohorts: the area under the curve (AUC) was 0.987 in the training set, and 0.892, 0.884, and 0.851 in the internal validation set and two external test sets (Test 1, Test 2), respectively. Conclusion The integration of deep learning, radiomics features derived from abdominal ultrasonography, and clinical data demonstrates distinct advantages in predicting 48-hour early intussusception recurrence (EIR). It can provide robust support for clinicians to formulate individualized post-reduction management strategies.