A novel MRI-based deep learning-radiomics framework for evaluating cerebrospinal fluid signal in central nervous system infection.

Introduction: Accurate and timely diagnosis of central nervous system infections (CNSIs) is critical, yet current gold-standard techniques like lumbar puncture (LP) remain invasive and prone to delay. This study proposes a novel noninvasive framework integrating handcrafted radiomic features and deep learning (DL) to identify cerebrospinal fluid (CSF) alterations on magnetic resonance imaging (MRI) in patients with acute CNSI. Methods: Fifty-two patients diagnosed with acute CNSI who underwent LP and brain MRI within 48 h of hospital admission were retrospectively analyzed alongside 52 control subjects with normal neurological findings. CSF-related signals were segmented from the ventricular system and sub-lentiform nucleus parenchyma, including perivascular spaces (PVSs), using semi-automated methods on axial T2-weighted images. Two hybrid models (DenseASPP-RadFusion and MobileASPP-RadFusion), fusing radiomics and DL features, were developed and benchmarked against base DL architectures (DenseNet-201 and MobileNet-V3Large) via 5-fold nested cross-validation. Radiomics features were extracted from both original and Laplacian of Gaussian-filtered MRI data. Results: In the sub-lentiform nucleus parenchyma, the hybrid DenseASPP-RadFusion model achieved superior classification performance (accuracy: 78.57 ± 4.76%, precision: 84.09 ± 3.31%, F1-score: 76.12 ± 6.86%), outperforming its corresponding base models. Performance was notably lower in ventricular system analyses across all models. Radiomics features derived from fine-scale filtered images exhibited the highest discriminatory power. A strict, clinically motivated patient-wise classification strategy confirmed the sub-lentiform nucleus region as the most reliable anatomical target for distinguishing infected from non-infected CSF. Discussion: This study introduces a robust and interpretable MRI-based deep learning-radiomics pipeline for CNSI classification, with promising diagnostic potential. The proposed framework may offer a noninvasive alternative to LP in selected cases, particularly by leveraging CSF signal alterations in PVS-adjacent parenchymal regions. These findings establish a foundation for future multicenter validation and integration into clinical workflows.