Abstract:
The fast detection of spinal fractures is crucial for preventing neurological deterioration or paralysis [1]. However, detecting fractures on scans is challenging, especially for elderly individuals due to weak bones. In recent years, artificial intelligence (AI) has significantly improved healthcare outcomes by detecting cancers early and distinguishing between benign and malignant tumors. This thesis aims to employ computer vision for the detection of cervical spine fractures using computed tomography (CT) scans. The developed model classifies CT scans into several classes corresponding to fractures at specific cervical vertebrae (C1 to C7) or no fracture. Initial considerations involved using three-dimensional models to analyze CT scans, but due to limited computing resources, the project shifted focus to two-dimensional models despite the restricted access and limited annotations typical of medical images. The resulting algorithm utilizes two-dimensional convolutional neural networks to perform sequential classifications, first identifying the presence of a fracture on a CT slice and subsequently determining the specific vertebra involved. Despite challenges, the second classifier achieved a high accuracy of 97% in tests. However, the comprehensive analysis of prediction results was not implemented due to time constraints. This abstract succinctly outlines the objectives, methods, and outcomes of the thesis, emphasizing the transition from three-dimensional to two-dimensional analysis due to resource constraints and achieving notable accuracy with the chosen method.
