Serum identification of childhood brain tumors by fused laser-induced breakdown spectroscopy and Raman spectroscopy.
In serum from 12 children with brain tumors and 10 healthy children, fused LIBS and Raman spectral features classified tumor status with a mean accuracy of 92.29% across repeated participant-level train/test splits.
Open original publication →What the AI sees
In serum from 12 children with brain tumors and 10 healthy children, fused LIBS and Raman spectral features classified tumor status with a mean accuracy of 92.29% across repeated participant-level train/test splits.
Research significance
The reported evidence supports preliminary feasibility for serum-based tumor identification; it may, by inference, become a minimally invasive adjunct for diagnosis or monitoring that could inform treatment decisions, but no treatment-selection, outcome, or longitudinal-monitoring benefit was tested.
Source abstract
Childhood brain tumors (CBTs) are common central nervous system tumors in children, whereas conventional diagnostic methods are costly, time-consuming, invasive, and unsuitable for frequent monitoring. Here, we explored the feasibility of integrating laser-induced breakdown spectroscopy (LIBS) and Raman spectroscopy (RS) for CBT identification using dried serum because these techniques enable rapid analysis, require simple sample pretreatment, and are minimally destructive or non-destructive. Serum samples were collected from 10 healthy children and 12 patients with CBTs. Participant-level divisions of the training and test sets (DTTs) were performed at an approximately 7:3 ratio and repeated 10 times to assess the robustness of classification performance across different cohort compositions. Laser-etched silicon wafers (LSWs) with different mesh sizes were prepared as serum substrates. After optimization of the mesh parameters, the relative standard deviations of the spectral features decreased by 1.74%-73.16% for LIBS and 23.46%-25.32% for RS. An ensemble classification method, termed support vector machine (SVM)+convolutional neural network (CNN)-random forest (RF), was proposed to improve classification accuracy. Compared with the individual SVM, CNN and RF classifiers, the SVM+CNN-RF method using selected spectral features achieved the best single-modality performance, with mean accuracies across the 10 DTTs of 86.57% for LIBS and 86.29% for RS. After fusion of the LIBS and Raman features followed by SVM+CNN-RF classification, the accuracies ranged from 90.29% to 94.29%, with a mean accuracy of 92.29%, representing improvements of 5.72% and 6.00% over LIBS and RS alone, respectively. Overall, these results demonstrated that LIBS-Raman fusion combined with the SVM+CNN-RF method was a feasible strategy for identifying CBTs from dried serum and showed promise as an adjunctive approach for clinical diagnosis.