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

Abstract

Background: Early recurrence of intussusception frequently occurs within 48 h after reduction, yet current assessments lack objective tools for accurate risk prediction.

Objective: This multicenter study sought to develop and validate a predictive model for early intussusception recurrence using abdominal ultrasound imaging.

Materials and methods: We retrospectively collected data from 1,314 primary pediatric intussusception cases across three hospitals (2016-2024). Cases from center 1 were split into training (60%) and internal validation (40%) sets, and those from centers 2 and 3 composed the external test sets. An automated segmentation model (DeepLabv3 with ResNet101) was developed for lesion delineation. Radiomics and deep learning features were extracted from the segmented regions. Three classifiers (K-nearest neighbors, random forest, and extreme gradient boosting) were evaluated on the radiomics, deep learning, and fused feature sets. We integrated the fused model with clinical variables to construct a nomogram.

Results: The Dice coefficients of the segmentation model were 0.935 (seg-training) and 0.862 (seg-test). The key clinical predictors were age, onset-to-presentation time, and vomiting. After seven radiomics features and nine deep learning features were selected, we integrated the fused model with clinical variables to construct a nomogram. It showed good discrimination (internal validation area under the receiver operating characteristic curve (AUC), 0.892; external test set AUCs, 0.884 and 0.851, respectively). The results of the calibration and decision curve analyses supported its potential as a risk stratification tool.

Conclusion: Integrating ultrasound-derived deep learning, radiomics, and clinical data shows promise for predicting recurrence of primary intussusception within 48 h. The model may aid early risk stratification, although prospective validation in diverse populations is needed.