Evaluation of U-Net model in the detection of cervical spine fractures

Abstract

The cervical spine is composed of seven vertebrae from C1 to C7 with a lordotic curve (C-shaped curve) and joints between vertebrae for spine mobility. A computed tomography (CT) is commonly used by experts and physicians in imaging diagnosis to give information about the cervical spine and vertebrae in the neck. Diseases such as spinal stenosis (narrowing of the spinal canal), herniated discs, tumors, and fractures in the cervical spine can be diagnosed by CT scans. Quickly detecting the presence, and location of cervical spine fractures in CT scans helps physicians prevent neurologic deterioration and paralysis after trauma. Throughout this thesis, a U-Net model was trained for semantic segmentation on approximately 2019 study instances with provided CT images, while only 87 of them have been segmented by spine radiology specialists. After that, a combination of 2D CNN and bidirectional GRU deep learning models was used for the detection of fractures in each vertebra, as a classification task. The objectives of this research are to develop two deep-learning models for detecting and localizing cervical spine fractures and evaluate the ongoing research activities on semantic segmentation and classification in the medical field. This research aims to use a semantic segmentation algorithm in deep learning by using U-Net architecture to estimate the location of each cervical vertebra, as well as propose a deep convolutional neural network (DCNN) with a bidirectional GRU memory (Bi-GRU) layer for the automated detection of cervical spine fractures in CT images. This approach was trained and tested on a dataset provided by RSNA (a team of the American Society of Neuroradiology and Spine Radiology). Furthermore, the critical factors, such as preprocessing techniques and specialized loss functions were explored that must be taken into consideration when segmenting 3D medical images. Whether used as a standalone framework for segmentation and classification tasks or as an integrated backbone for medical image processing, this architecture is flexible enough to accommodate other models. The proposed approach yields results that are comparable to those of existing techniques, but it can be improved by using larger image sizes and more advanced GPU workstations that will reduce the overall processing time. Future research will be using other pretrained networks as an encoder and increasing image sizes to examin the performance improvmet of the architecture which needed more advanced computational resources and also integrate the current architecture into a simulated crash scenarios to use in various applications such as producing protecive sport equipments.