Machine Learning Integration of MRI Intratumoral and Peritumoral Radiomics Features for Predicting PNSTs Postoperative Complications.
Authors
Affiliations (2)
Affiliations (2)
- Wound Repair Department, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China (J.H.L, F.Y.T., J.F.W., Y.H.C., R.W.H., J.W.); Hand Surgery & Peripheral Nerve Surgery, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China (J.H.L, F.Y.T., J.F.W., Y.H.C., R.W.H., J.W.).
- Wound Repair Department, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China (J.H.L, F.Y.T., J.F.W., Y.H.C., R.W.H., J.W.); Hand Surgery & Peripheral Nerve Surgery, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China (J.H.L, F.Y.T., J.F.W., Y.H.C., R.W.H., J.W.). Electronic address: [email protected].
Abstract
Postoperative complications (15-76%) substantially affect patients with peripheral nerve sheath tumors (PNSTs), yet objective preoperative risk tools are lacking. Magnetic resonance imaging (MRI) radiomics has focused mainly on intratumoral regions, while the predictive value of the peritumoral microenvironment remains unclear. In this retrospective single-center study (Dec 2015-Jan 2024), 280 pathologically confirmed PNST patients with preoperative MRI were randomly split 8:2 into training (n=224) and test (n=56) cohorts. Intratumoral (Intra) and Peritumoral (Per; 2Â mm expansion) regions were manually segmented and 1197 radiomic features extracted. Four models were constructed: Intra-model, Per-model, a fused Intra-model+Per-model region model (Imagefusion) and a concatenated Intra-model+Per-model feature model (intraPeri2mm). After reliability filtering, t-tests and least absolute shrinkage and selection operator selection, models were trained with machine-learning classifiers. Clinical predictors were assessed, and model performance evaluated using area under the receiver operating characteristic curve (AUC) and decision-curve analysis (DCA). Diabetes was the only independent clinical predictor, and the clinical model achieved a test AUC of 0.599. In the test cohort, both fusion models outperformed single-region models (intraPeri2mm AUC 0.899; Imagefusion AUC 0.895), with consistently greater net benefit on DCA. Incorporating diabetes yielded a small, nonsignificant gain for Imagefusion (Combined_1 AUC 0.917) and no further improvement for intraPeri2mm (Combined_2 AUC 0.889). All radiomics models significantly exceeded the clinical-only model (all p<0.001). Integrating intra- and peritumoral radiomics enables effective preoperative prediction of PNST postoperative complications. An Imagefusion +clinical pathway offers robust clinical net benefit when clinical data are standardized, whereas an intraPeri2mm -only strategy may be preferable where clinical data are limited.