Label-informed Data Augmentation for DBS Optimization: Synthesizing fMRI Maps with SPADE-VAE Network to Improve DBS Parameter Classification

Motivation: Optimizing deep brain stimulation (DBS) parameters for Parkinson's disease (PD) has a high time to optimization per patient (TPP). Functional MRI (fMRI) and deep learning (DL)-based optimization can substantially reduce TPP, but faces data scarcity and class imbalance issues. Goal(s): To improve DBS optimization by addressing DBS-fMRI data limitation and imbalance issues. Approach: A SPADE-VAE model was trained to synthesize realistic optimal DBS-fMRI maps at varying optimization levels by blending latent representations of acquired optimal and non-optimal DBS-fMRI responses. Results: The trained SPADE-VAE model generated realistic synthetic DBS-fMRI response maps, which improved the AE-MLP DBS parameter classification accuracy from 80% to 91%. Impact: By producing realistic synthetic DBS-fMRI maps, the trained SPADE-VAE model addresses data gap issues in DL-based fMRI-DBS optimization, corrects class imbalances, enhances classification accuracy and ultimately reduces the time to optimization per patient in Parkinson's disease treatment.