| Summary: | 碩士 === 國立臺北科技大學 === 電機工程研究所 === 107 === Lumbar spines are the largest segments of the spinal canal, they also involve nerve roots enclosed by dura sac and cerebrospinal fluid (CSF). In neurosurgery, lumbar spinal stenosis (LSS) is the leading preoperative diagnosis caused by bones degeneration for adults older than 65 years. When the spinal nerves in the lower back are choked, it could lead to the spinal canal stenosis. The symptoms of LSS have a significant impact on essential contents within the spinal canal, especially reduction of CSF.
In clinical practice, severity of LSS is usually evaluated on the lumbar axial T2 MRI and is often referred to as the cross section area (CSA) of the spine with clinical neurological symptoms. Generally, measurement of CSA relies on experienced neurosurgeons to mark the area of stenosis. The process spends time and is difficult to quantify the LSS. With vast applications of machine learning and image processing, literatures exploited to use SVM classifiers to measure the volume of CSF. However, advantages of SVM cannot work in the cases with highest severity because of absent of training samples. On the other hand, performance of classifiers would be affected since MRI can only provide T1 and T2 sequences as features.
This manuscript extended state-of-the-arts (deep learning architecture: U-Net) and proposed two major methods to improve CSF segmentation performance. To apply concepts of random process, band expansion process considered the original image as random variables. By different correlations and nonlinear statistical functions among random variables, band expansion process generates new band images to fill vacant sequences for enhancement of performance. Another algorithm is an unsupervised thresholding framework. This method was derived from moment generating functions to simulate the unknown distribution of T1 and T2 luminance, and segmented CSF area with an adaptive dynamic threshold. This proposed algorithm effectively reduced classification error and improved performance of classifiers especially on higher severity of stenosis.
Several methodologies were included in the experimental studies for comparison: features expansion by band expansion process, segmentation by U-Net in deep learning, and a proposed thresholding framework based on statistical indicators. Advantages and disadvantages of these algorithms would be further evaluated and covered in the following sections.
|
|---|
