腦部結構差異量化分析之區塊多變量型態法

Abstract

磁振造影已經成為主要的醫療成像技術，用於了解腦部或是身體上的結構以及功能，它不僅廣泛應用於臨床診斷，而且在神經影像學研究。近年來，以體素為基礎的形態計量學 (voxel-based morphometry, VBM) 最常被使用於研究分析群體間大腦結構的差異性，以比較每一個體素的方式來找出其差異性。 VBM 以統計的方式量化分析群組間的差異性，這個以體素為基礎的分析方法有能力上的限制，導致它無法偵測群組織間細微的變化。根據我們的經驗指出，同時地將所有的體素一起做分析會使區域間不相關的體素相互影響，這樣的分析方法是不恰當的。

本研究中我們提出了一個新的腦部結構分析方法，以區域為單位的多變量形態計量學方法，可用於偵測群組間腦部結構的差異性。與體素為基礎的分析方法相比較之下，我們提出的方法是同時考慮一個區域下所有的體素以多變量的方式來分析群組間腦部結構的差異性，而最重要的是我們將腦部結構區分成許多個小區域，藉此達到在相同區域下的所有體素有相同的關聯性。多變量的方法是採用線性鑑別度分析 (linear discriminant analysis, LDA) 是用來找出對於群組間結構差異性最具鑑別力的投影軸。位於該投影軸上的每一元素代表相對映體素具有的差異性鑑別能力。此鑑別能力可以視為用來評估影像群組間每一體素結構差異之程度等級 (significance level)。

我們的實驗方法包含了兩個部分，第一個部分是透過模擬一個區域的萎縮，用來驗證以區域為單位的多變量型態計量法的效能，另一個是應用在重性抑鬱障礙 (Major depressive disorder) 以及雙極性情感疾病 (Bipolar disorder)的腦部結構分析。透過比較可以了解，以區域為單位的多變量型態計量法確實可偵測到群組間較細微的搞部結構差異。而此方法也比以體素為基礎的形態計量法更明顯地找出和病理上相關的腦部結構差異。

總結，我們提出了一個新的腦部結構分析方法，這個方法是以區域為單位的分析群組間的差異，而且所採用的多變量方法不會有資訊上的遺失，在分析時能更用更多的資訊更有效地分析腦部結構細微的差異性。

Magnetic resonance imaging (MRI) has become primarily a medical imaging technique to visualize the structure and function of the body or brain. It is widely used not only in clinical diagnosis but also in neuroimaging research. In recent years, voxel-based morphometry (VBM) is one of the most popular technique for the analysis of structural brain discrepancy between different subject groups, in a voxel-wise manner. VBM analysis detects group differences by voxel-wise statistics comparisons which have limited power to identify subtle differences between two populations. And according to our experience, we figure out that when dealing with all features simultaneously, features in different regions of whole brain may be unrelated with each other it is incorrect that we take all features into consideration at one time.

In this work, we propose a parcellation-based multivariate morphometry method which can be used to detect the anatomical discrepancy in brain between two groups. Compared to the voxel-wise manner in VBM, the proposed method detects brain discrepancy in a multivariate manner by simultaneously taking all voxels within an area (or a region) in consideration. The most important idea is that we divide brain into several parts when analyzing such that all features in the same region may be correlated with each other. Linear discriminant analysis (LDA) is used to determine the most discriminant projection vector, also called a discriminant map separating two populations. Each parameter of the most discriminant vector represents the discrimination level of each voxel. That is, based on the discriminant map, each parameter stands for a significant level with each voxel.

To demonstrate the performance of the parcellation-based multivariate method, we carried out experiments by using the simulation data set and on a real medical data composed of MRI of subjects with major depressive disorder (MDD) and bipolar disorder (BD). The results with simulation data analysis have shown that the parcellation-based multivariate method has a better performance than VBM from the area under the ROC curve by comparing to VBM method. The results with real data analysis have also shown that our proposed method reveals several important findings.

In conclusion, we have proposed a parcellation-based multivariate method for characterizing group differences. And all voxels within the same region are simultaneously taken into consideration. Moreover, our proposed method uses no feature reduction within analysis thus no information is lost. Our proposed method have shown that the parcellation-based multivariate morphometry analysis has a good performance on subtle and widely-distributed structural difference and it is more flexible within analysis.