A deep learning-based approach for multi-station spine MRI stitching

MRI is widely used for the diagnostic evaluation of the spine, but the anatomical length of the vertebral column often exceeds the high-quality FOV achievable in a single acquisition. Therefore, whole-spine imaging commonly requires multi-station protocols, where partially overlapping regions are acquired sequentially and subsequently combined into an extended-FOV representation. This stitching process is technically challenging because adjacent stations may present patient motion, limited overlap, intensity inhomogeneity and geometric distortion near the boundaries of the acquisition volume. This dissertation investigates a DL-based framework for the automated stitching of multi-station spinal MRI images: the implemented pipeline includes dataset creation, affine regression, expanded-canvas stitching, SimpleITK-based refinement and quantitative evaluation using the DISTS, SSIM and PSNR metrics.The training of the neural network, based on global affine parameter regression, was performed using synthetically generated image pairs obtained from spinal MRI images, where known 2D affine transformation parameters provided supervised ground-truth transformation parameters. The model was first trained and tested on sagittal spinal MRI images and then fine-tuned on coronal images due to the imbalance of the available public dataset. The results show that the proposed affine regression strategy represents a promising baseline for anatomically conservative spinal MRI stitching; however, visual inspection also revealed residual failure cases, particularly in the coronal view, confirming that the current implementation is not yet suitable for clinical or commercial deployment. The main limitations are related to the lack of public complete multi-station whole-spine MRI datasets and to the limited availability of coronal images. Future work should therefore focus on whole-spine dataset acquisition, vertebra-aware constraints and more extensive clinical validation.