Early Diagnosis of Neuropsychiatric Systemic Lupus Erythematosus by Deep Learning Enhanced Magnetic Resonance Spectroscopy

Abstract

The neuropsychiatric systemic lupus erythematosus (NPSLE) has higher disability and mortality rates, which is one of the main causes of death in systemic lupus erythematosus (SLE) patients. Magnetic resonance spectroscopy (MRS) can detect the changes of metabolites in different intracranial areas in vivo in patients with SLE, so as to provide evidence for the early diagnosis of NPSLE. Different from the conventional single-voxel MRS, which can only screen one brain region with one metabolic change, we simultaneously detect 13 kinds of intracranial metabolic changes in nine brain regions by multivoxel proton MRS (MVS). We use a recursive feature elimination algorithm to select the most related metabolites for better identifying NPSLE. To accurately diagnosis NPSLE by these intracranial metabolites, we train a support vector machine deep stacked network (SVM-DSN) for quantitative analysis of these metabolites. Comparing with the conventional statistic method, which is about 70% of accuracy, the proposed model achieves 97.5% of accuracy for NPSLE diagnosis. We conclude the trained SVM-DSN can effectively analyze the metabolites obtained by multivoxel proton MRS for NPSLE diagnosis, which may help to early diagnosis and intervention of NPSLE, and alleviate the bias of manual screening.