Cerebrospinal fluid lumen segmentation with artificial intelligence techniques

Analysis of the complex behaviors of Cerebrospinal Fluid (CSF) has become essential for disease diagnosis since an abnormal circulation of CSF leads to the development of illnesses. An innovative non-invasive technique called Phase-Contrast Magnetic Resonance Imaging (PC-MRI) allows both time series measurements of CSF flow in the cardiac cycle and the region of interest's (ROI) visual data. The technique requires a significant amount of time and experience to determine the exact ROI location and perform the necessary examination and evaluation of the CSF flow. This dissertation suggests two automatic methods for segmenting the lower thoracic region to address that issue. First, machine learning techniques were implemented using pulsatile data as a feature. The second, three-dimensional (3D) U-Net model, fed with pulsatile data from PC-MRI as the third dimension for training, was designed. The dataset includes 2176 phase and rephase 3-tesla PC-MRI images from 57 slabs of 39 control subjects and individuals with idiopathic scoliosis. The 5-fold cross-validation procedure was used to evaluate the 3D Attention U-Net model after training and achieved an average weighted performance of 97% precision, 95% recall, 98% F1 score, and %95 area under curve. The model's success was also measured using the CSF flow waveform quantities. The mean and peak flow rates through the labeled and predicted CSF lumens had a significant correlation coefficient of 0.96 and 0.65 sequentially. From what is known, this thesis is the first fully automatic 3D deep learning implementation to segment CSF-containing spaces in the spinal using both spatial and pulsatile flow information in PC-MRI data. This work is expected to attract future research using PC-MRI pulsatile data for training deep models.