Abstract
Automated spinal structure segmentation in sagittal MRI remains a non-trivial task due to high inter-patient variability and ambiguous anatomical boundaries. We propose CAFR-Net, a Transformer-contrastive hybrid framework that jointly models global semantic relations and local anatomical priors to enable precise multi-class segmentation. The architecture integrates (1) a multi-scale Transformer encoder for long-range dependency modeling, (2) a Locally Adaptive Feature Recalibration (LAFR) module that reweights feature responses across spatial-channel dimensions, and (3) a Contrastive Learning-based Regularization (CLR) scheme enforcing pixel-level semantic alignment. Evaluated on the SpineMRI dataset, CAFR-Net achieves state-of-the-art performance, surpassing prior methods by a significant margin in Dice (92.04%), HD (3.52 mm), and mIoU (89.31%). These results underscore the framework's potential as a generalizable and reproducible solution for clinical-grade spinal image analysis.