Magnetic Resonance Imaging-Based Molecular Inference and Cerebrospinal Fluid-Derived Cell-Free DNA Profiling in Central Nervous System Tumors.
Authors
Affiliations (6)
Affiliations (6)
- Department of Physiology, Ajou University School of Medicine, Suwon, Korea.
- Department of Computer Education, Gyeongin National University of Education, Incheon, Korea.
- Department of Neurosurgery, Soonchunhyang University Seoul Hospital, Seoul, Korea.
- Department of Laboratory Medicine, Soonchunhyang University Seoul Hospital, Seoul, Korea.
- Department of Laboratory Medicine, Kangbuk Samsung Hospital, Seoul, Korea.
- Department of Neurosurgery, Soonchunhyang University Cheonan Hospital, Cheonan, Korea.
Abstract
Accurate molecular characterization is increasingly important in the management of central nervous system (CNS) tumors; however, tissue-based molecular testing remains invasive and constrained by sampling limitations. We investigated the feasibility of cerebrospinal fluid (CSF)-derived cell-free DNA (cfDNA) profiling together with magnetic resonance imaging (MRI)-derived phenotype inference as complementary approaches for molecular interpretation in CNS tumors. Targeted next-generation sequencing was performed on tumor tissue genomic DNA and matched CSF cfDNA. A multisequence MRI-based deep learning framework was established to estimate tumor phenotype probabilities from T1-weighted, T2-weighted, fluid-attenuated inversion recovery, and contrast-enhanced T1-weighted images. Gene-level molecular tendencies were inferred using a phenotype-conditioned probabilistic framework combining imaging-derived phenotype probabilities with class-associated variant frequencies. Among 91 collected CSF samples, 10 met the sequencing input criterion for cfDNA analysis. Within this limited subset, CSF-derived cfDNA showed concordance with matched tumor tissue for tier I and II variants, while additional source-specific variants were also observed. MRI-derived phenotype probabilities demonstrated modality-dependent discrimination between glial and nonglial tumors, with the highest performance observed for T2-weighted MRI (area under the curve, 0.833). Phenotype-conditioned molecular inference also showed an overall tendency toward concordance with sequencing profiles, supporting the feasibility of deriving complementary molecular information from imaging-based phenotype estimates. MRI-derived phenotype information and limited CSF cfDNA profiling may provide complementary molecular information on CNS tumors. The proposed framework is exploratory, does not constitute direct variant prediction, and is not intended to replace tissue-based molecular testing, particularly in settings where conventional molecular profiling remains constrained.