Deep learning-derived functional MRI connectivity maps of cervical spinal cord segments during hand grasping

Task-based fMRI of the cervical spinal cord is limited by physiological noise and small sample sizes. The primary objective of this study was to examine whether a convolutional neural network (CNN)-based denoising framework can improve the detection of task-dependent functional connectivity between cervical spinal segments using generalized psychophysiological interaction (gPPI) analysis. An open dataset of 26 healthy adults performing a hand-grasp task with alternating right- and left-hand blocks during 3 T fMRI was analyzed. A CNN with Inception V3 transfer learning (91.0% slice-level and 89.2% block-level accuracy) was trained to classify grasp versus rest, and the trained model was then utilized to derive denoised time series from masks of the spinal segments C5-C8. A gPPI model with a C6 ventral horn seed was evaluated using FDR correction and permutation testing (n = 1000). C6 showed positive task-dependent connectivity with C7 (β = 0.72, p = 0.003) and C8 (β = 0.58, p = 0.018), and negative connectivity with C5 (β = -0.31, p = 0.041), which is broadly consistent with the organization of the cervical myotomes. The C6-C5 effect was marginal and was not detected when conventional time series were used (p = 0.21), and therefore it requires replication. Grad-CAM maps aligned with ventral horn motor pools (Dice = 0.89 at C7). These results suggest that deep learning-based denoising may increase the sensitivity of spinal cord fMRI to segmental connectivity, although validation in larger cohorts is needed.