Accurate, subject-specific estimates of spinal loading enhance our understanding of the mechanisms underlying back pain and degeneration. However, established biomechanical analyses require labor-intensive musculoskeletal model generation and computationally expensive simulations, limiting their scalability and clinical applicability. We present a deep-learning surrogate model that predicts spinal loading directly from imaging-derived anatomy, aiming to replace time-consuming subject-specific simulations. We used our automated pipelines for segmentation and for subject-specific musculoskeletal torso modeling and simulation from full-torso medical images [1] to generate reference spinal loading during upright standing. Two datasets were created based on: (A) inverse-dynamics (ID) simulations without muscle contributions to estimate net joint loads from MRI of a NAKO (German National Cohort) subcohort (n = 860), and (B) combined ID with static optimization incorporating muscle influence for physiological spinal load and muscle force estimation from clinical CT (n = 760). For load prediction, we used a 3D convolutional neural network with a ResNet3D backbone [2]. Multi-channel inputs were formed by concatenating 3D volumes with spine and tissue segmentation masks, and identical train/validation/test splits (80/10/10) were used. On (A), the network regressed net shear and compression forces and moments, achieving strong accuracy for compression (RMSE ≈ 16.3 N, NRMSESD ≈ 0.22) and moderate accuracy for anteroposterior shear (RMSE ≈ 35.2 N, NRMSESD ≈ 0.46) and flexion moments (RMSE ≈ 3.1 Nm, NRMSESD ≈ 0.56). On (B), a unified multi-task model jointly predicted shear, compression, and muscle forces. Muscle fascicle forces reached RMSE ≈ 1.7 N (NRMSESD ≈ 0.46), and load prediction was comparable to (A). Average inference time was ~1.3 s per sample, compared to >2.5 min for automated modeling and simulation. Overall, the proposed simulation-trained neural surrogate represents a step towards clinically feasible spinal load assessment by enabling rapid, anatomy-driven load estimation from both research-cohort and routine clinical imaging. Further experiments and hyperparameter tuning are necessary to reliably predict comprehensive spinal biomechanics from imaging data.