Magnetic Resonance Spectroscopy Deep Learning with Magnetic Resonance Background Generator Enables In Vivo Metabolite...

Short-TE Proton Magnetic Resonance Spectroscopy (SPMRS) allows non-invasive, radiation-free detection of biomolecules including key brain metabolic biomarkers for neurological diseases. However, SPMRS also contains signals from macromolecules, lipids and baseline. Existing Deep Learning (DL) quantification methods for SPMRS overlook baseline interference.

To address this, we proposed a novel DL quantification incorporating macromolecules, lipids and baseline (called background signal). First, background signal generators were built by capturing feature distributions of 200 in vivo background signals. Then, the synthetic SPMRS spectra were generated, composed of synthetic metabolite components based on the quantum mechanical and exponential models, background signal generated by the background generator and noise. Finally, the synthetic SPMRS were used to train the advanced DL quantification model QNet (our earlier study).

The proposed DL method was validated on a large-scale healthy human brain dataset (BigGABA) and our collected data from different manufacturers (totaling 216 spectra). Furthermore, it was applied to classification of minimal hepatic encephalopathy (totaling 176 spectra), highlighting its clinical potential.

We proposed and validated a novel deep learning-based quantification method for SPMRS which is explicitly taken into account the background effect.

The proposed method is incorporated the background effect into the DL framework of SPMRS quantification, and DL quantification of SPMRS is utilized for the clinical diagnosis of minimal hepatic encephalopathy.