Radiomics and machine learning for differentiating pediatric intramedullary spinal tumors: a pilot study.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
OBJECTIVES: Conventional magnetic resonance imaging faces limitations in differentiating among pediatric intramedullary spinal tumors (IMSTs). We evaluated radiomics-based machine learning (ML) models using sagittal T2-weighted images to differentiate pilocytic astrocytomas (PA) and ependymomas (EP), independently, from other IMSTs. MATERIALS AND METHODS: This single-center retrospective study included pediatric patients diagnosed with IMST from 2000 to 2023. Manual tumor segmentation was performed on sagittal T2-WI, in correlation with sagittal contrast-enhanced T1-weighted images using ITK-SNAP. Radiomic features were extracted using PyRadiomics. Effect size validation assessed feature analytical suitability prior to ML implementation. Three ML models with two feature selection methodologies were deployed using nested leave-one-out cross-validation. Balanced accuracy, Matthews correlation coefficient (MCC), sensitivity, specificity, positive/negative predictive values (PPV, NPV), F1 score, area under the receiver operating characteristic curve (AUC), Youden's J, and Cohen's d were calculated. RESULTS: Forty patients, aged 9.2 ± 5.4 years (mean ± standard deviation), 21 females, were analyzed. Tumors comprised PAs (n = 16), EPs (n = 11), gangliogliomas (n = 6), and others (n = 7). Cohen's d was 0.356 for PAs and 0.741 for EPs. PA radiomics showed inadequate capacity, resulting in suboptimal ML performance. Random forest with least absolute shrinkage and selection operator (LASSO) achieved superior EP classification using an average of four features (Youden's J = 0.243): balanced accuracy 0.834 (95% confidence interval: 0.723-0.955, p < 0.0001), AUC 0.838, sensitivity 0.909, specificity 0.759, PPV 0.588, NPV 0.957, and MCC 0.603. CONCLUSION: Radiomics-based ML differentiated EPs but not PAs from other pediatric IMSTs using sagittal T2-weighted images. KEY POINTS: Question Can radiomic-based machine learning models differentiate pediatric intramedullary spinal tumors? Findings Random forest with LASSO feature selection achieved acceptable performance in distinguishing ependymomas from other IMSTs, with balanced accuracy of 0.834, AUC of 0.838, and MCC of 0.603. Relevance statement This review highlights the emerging role of time-resolved ("4D") approaches in postmortem imaging, demonstrating how integrating temporal information can enhance reconstruction of injury mechanisms, postmortem processes, and medicolegal interpretation, thereby expanding forensic radiology beyond static documentation toward process-oriented forensic investigation.