Abstract
Multimodality imaging is an emerging research topic in neuro-oncology for its potential of being able to demonstrate tumours in a more comprehensive manner. Diffusion-weighted magnetic resonance imaging (dMRI) and proton magnetic resonance spectroscopy ((1)H-MRS) allow inferring tissue cellularity and biochemical properties, respectively. Combining dMRI and (1)H-MRS may provide more accurate diagnosis for paediatric brain tumours than only one modality. This retrospective study collected 1.5-T clinical (1)H-MRS and dMRI from 32 patients to assess paediatric brain tumour classification with combined dMRI and (1)H-MRS. Specifically, spectral noise of (1)H-MRS was suppressed before calculating metabolite concentrations. Extracted radiomic features were apparent diffusion coefficient (ADC) histogram features through dMRI and metabolite concentrations through (1)H-MRS. These features were put together and then ranked according to the multiclass area under the curve (mAUC) and selected for tumour classification through machine learning. Tumours were precisely typed by combining noise-suppressed (1)H-MRS and dMRI, and the cross-validated accuracy was improved to be 100% according to naïve Bayes. The finally selected radiomic biomarkers, which showed the highest diagnostic ability, were ADC fifth percentile (mAUC = 0.970), myo-inositol (mAUC = 0.952), combined glutamate and glutamine (mAUC = 0.853), total creatine (mAUC = 0.837) and glycine (mAUC = 0.815). The study indicates combining MR imaging and spectroscopy can provide better diagnostic performance than single-modal imaging.