A Fully Automated Pipeline for Vertebral Structural Assessment From Medical Images. Application Under Metastatic Conditions.

Abstract

Spinal bone metastases often lead to vertebral fractures and other skeletal events that severely affect patients' quality of life. Predicting structural failure is essential for guiding treatment and preventing complications. However, conventional assessment tools have limited predictive power, highlighting the need for computational methods capable of simulating disease progression and its mechanical consequences. This study aimed to develop a fully automated, patient-specific methodology to predict vertebral structural behavior from computed tomography (CT) data, supporting clinical decision-making and treatment planning. The proposed pipeline integrates three main components: a deep neural network for semantic segmentation of vertebrae and metastatic regions from CT scans, the Coherent Point Drift (CPD) algorithm to ensure consistent alignment and automated definition of boundary conditions across datasets, and the Cartesian Grid Finite Element Method (cgFEM) to simulate the vertebral mechanical response under metastatic involvement. Model performance was evaluated by comparing the predicted outcomes with reference data, using precision, sensitivity, and specificity to assess reliability. The proposed workflow achieved full automation from CT imaging to fracture risk estimation. The segmentation module showed high accuracy across multiple metrics, enabling robust model generation. Geometric normalization and CPD-based boundary condition assignment standardized vertebral geometries across studies, while cgFEM simulations provided clinically interpretable metrics such as safety factors and stability variations associated with tumor size, location, and density. These analyses enabled the identification of scenarios linked to a higher fracture risk. The main limitations include the use of fixed boundary conditions and a predefined voxel threshold, which may reduce physiological realism and generate false positives. Overall, this work presents an end-to-end, patient-specific framework for automated fracture risk evaluation in metastatic vertebrae. By combining deep-learning-based segmentation, geometric normalization, CPD alignment, and cgFEM simulations, the method produces clinically relevant outputs that can guide therapeutic strategies. Future developments will focus on integrating patient-specific loading data and predictive modeling to support incorporation into clinical decision support systems.