On the biomechanical side, we have conducted a series of tasks and research efforts aimed at advancing the understanding and modeling of spinal biomechanics through patient-specific musculoskeletal models and finite element models (FEM). Our primary objective was to validate an automatically generated patient-specific musculoskeletal model of the spine, which simulates static loading tasks using CT and MRI imaging data and a deep learning-based segmentation pipeline. We found that our model's results closely correlated with in vivo data from the literature, indicating that our process is suitable for systematic investigation of spinal biomechanics in large datasets. We also investigated the effects of different degrees of individualization on observed effects of morphological characteristics on lumbar loads by creating 279 models from CT data of 93 patients. We discovered that while significant effects of morphological parameters on lumbar loads were identified, the strength of these effects decreased with the inclusion of more individual characteristics. Further, we compared different methods for fiber reinforcement in the annulus fibrosus of a human intervertebral disc model, finding that the Holzapfel-Gasser-Ogden (HGO) model provided the highest accuracy in simulations.
To streamline the modeling process, we developed an automated pipeline for biomechanical simulations of vertebrae and intervertebral discs, using automated MRI segmentations. This pipeline successfully generated and simulated patient-specific models with high bio-fidelity, robustness, and time-efficiency. Our conference contributions included a pipeline for the automated generation of individualized musculoskeletal spine models, highlighting differences in spinal loading related to curvature in large patient cohorts. We also explored the role of spinal muscles in lumbar loads during static loading tasks and conducted a multivariant analysis on the effects of six morphological parameters on lumbar loading, identifying significant correlations with considerable variability yet to be explained.
In computer vision, we addressed the challenge of accurately delineating posterior spine structures in MRI. We translated T1-weighted and T2-weighted images into CT images from 263 pairs of CT/MR series using landmark-based registration. We compared various 2D and 3D image translation methods, including Pix2Pix and DDIM, evaluating the results with peak signal-to-noise ratio and Dice similarity coefficients. A publicly available segmentation network was used to segment the synthesized CT datasets. Based on this work, we developed SPINEPS, an open-source deep learning segmentation framework designed to accurately segment 14 essential spine structures in T2-weighted MRI scans. Our approach involved training two models using a combination of CT and T2-weighted segmentations on datasets including 218 subjects from the SPIDER dataset and 1,449 subjects from the German National Cohort. We also generated synthetic sagittal T1-weighted fast spin echo (T1w FSE) and short tau inversion recovery (STIR) images from sagittal T2-weighted (T2w) FSE and axial T1w gradient echo Dixon sequences. Using datasets from the Study of Health in Pomerania (SHIP), the German National Cohort (NAKO), and an internal dataset, we employed 3D Pix2Pix deep learning models to enhance image quality. Additionally, we proposed an unpaired inpainting superresolution algorithm to address the limitations of anisotropic 2D axial or sagittal spine MRI images. By generating synthetic training pairs and modeling various MR acquisition challenges, we trained diffusion-based superresolution models, enabling the separation of individual vertebrae instances and improving automatic segmentation. Finally, we compared different methods for identifying the thoracolumbar junction, evaluating rib-based assessments against vertebral shape-based classifications. In addition, we developed a full-body tissue and organ segmentation model based on T1-w gradient echo Dixon images.
This gives us the possibility to reliably assess 30k spines in NAKO as well as 6k spines in SHIP in a highly reliable fashion. It will be the basis for our biomechanical analysis as well as an image-based analysis of back pain in the NAKO cohort. We also implemented first methods of feature analysis, including spinal fracture detection and analysis of intervertebral disc geometry and degeneration.
