MRI multi-sequence deep learning integration with clinical profiles for pediatric viral encephalitis diagnosis.

Authors

Abstract

Pediatric viral encephalitis is an acute central nervous system infection caused by various viruses, with diverse clinical manifestations and challenges in early diagnosis. The traditional diagnostic methods lack sufficient sensitivity and specificity, highlighting the urgent need for an efficient, accurate, and non-invasive early diagnostic tool to guide clinical decision-making. This study aims to develop a clinical-imaging fusion model for pediatric viral encephalitis (VE) leveraging clinical factors and advanced magnetic resonance imaging (MRI) features. The diagnostic efficacy and application value of this model are explored. A retrospective analysis was conducted on 525 pediatric patients diagnosed with encephalitis at our research center. Patients were categorized into a viral encephalitis group (VE group) and a non-viral encephalitis group (non-VE group) based on etiological test results. MRI images (T1-weighted, T2-weighted, and diffusion-weighted imaging) and clinical characteristics were collected from all participants. Univariate and multivariate Logistic regression analyses were performed on clinical data to identify independent associated factors for VE. We manually delineated the Three-dimensional (3D) region of interest (ROI) layer by layer using ITK-SNAP software.3D brain tissues from different MRI sequences were segmented as ROI. Convolutional neural networks (CNNs) were employed to extract deep features from these ROIs. Following the fusion of deep features across the three sequences, the Least Absolute Shrinkage and Selection Operator (LASSO) was applied for dimensionality reduction to identify highly relevant deep features for VE. These features were integrated with clinically independent associated factors to build a fusion model using machine learning algorithms. In an independent test set, the model's performance was evaluated using Receiver Operating Characteristic (ROC) curves, Area Under the Curve (AUC), accuracy, sensitivity, specificity, predictive scores, and Decision Curve Analysis (DCA). Univariate and multivariate analyses revealed fever [OR = 1.455 (95% CI 1.170-1.811)], white blood cell (WBC) [OR = 0.434 (95% CI 0.299-0.629)] and C-reactive protein (CRP) [OR = 0.143 (95% CI 0.099-0.207)] as independent associated factors for pediatric viral encephalitis (VE). Among 6 MRI deep feature classifiers, the Logistic Regression (LR) classifier demonstrated superior performance, with AUCs of 0.934 (95% CI 0.909-0.959) and 0.899 (95% CI 0.851-0.946) in the training and test sets, respectively. The corresponding accuracy rates were 0.856 and 0.823, specificities were 0.862 and 0.746, and sensitivities were 0.851 and 0.885. In the clinical-imaging fusion model, AUCs reached 0.985 (95% CI 0.976-0.993) and 0.934 (95% CI 0.898-0.970) in the training and test sets, respectively. The accuracy rates were 0.935 and 0.867, specificities were 0.954 and 0.803, and sensitivities were 0.921 and 0.920. Decision Curve Analysis (DCA) confirmed the clinical utility of the fusion diagnostic model. The fusion model integrating deep features from multiple MRI sequences with clinical characteristics exhibits high diagnostic performance in pediatric viral encephalitis, offering a novel diagnostic tool for clinical practice.

Tags

Topics