運用有限混合模型執行三維果蠅神經影像機率圖之建構

Abstract

此研究的主要目的是建構果蠅嗅覺神經的機率圖影像(機率圖)與開發一種對果蠅嗅覺腦區(觸角葉)中的有著不同功能的局部神經元的嗅覺神經影像的自動化影像分割演算法。 在同一張三維影像中，有著許多不同功能的局部神經元，並呈現不同顏色與形狀。我們使用非監督式學習(統計分群)的方法來建構機率圖和開發自動化色彩分割演算法。

一般來說，影像資料有很多冗餘信息。 因此，我們必須使用各種各樣的方法去除雜訊。由於果蠅嗅覺神經很細，因此通常單獨使用空間濾波結果將會破壞神經的空間連續性。而常用频域濾波可能減少破壞神經的連續性的機會，但是這樣方法需要大量的計算。由於這些問題，在這項研究中，我們運用了影像資料的特徵並且開發方法去除雜訊並且修建機率圖。

基於我們的Brainbow技術的知識，我們可以知道不同的神經會呈現不同的顏色，但是在得到圖像過程中，顏色在不同的通道將有混疊現象。 這使萃取顏色特徵更加困難。在這項研究中，我們開發顏色萃取方法與建構混合模型並且運用這些特徵修建機率圖和開發一種自動化統計分群演算法。

The main purpose of this study is to construct the probabilistic graph image (Probabilistic Maps) of Drosophila’s olfactory neurons and develop an automated image segmentation algorithm for the Drosophila olfactory neurons in the olfactory brain regions (Antennal Lobe) which has many different functions of Local Neurons. In the same 3-D image, there are many different functions of local neurons, and these functions appear in different colors and shapes. We use the unsupervised learning (statistical clustering) approach to construct the probabilistic maps and develop an automated color segmentation algorithm.

In general, the image data has a lot of redundant information. Thus, we have to use various methods to remove noise. Because fruit flies have thin olfactory nerves, so using spatial filtering alone will often result in breaking the continuity of the nerves. The commonly used frequency domain filtering can reduce the chance of breaking the continuity of the nerves, but such a method requires a large amount of computation. Because of these issues, in this study, we utilized the characteristics of the image data and developed a method to remove noise as well as construct the probabilistic maps.

Based on our knowledge of the Brainbow technology, we can know that different nerves appear in different colors, but in the process of obtaining the images, colors in different channels will have the phenomenon of cross talk. This makes extracting the color features much more difficult. In this study, we develop some color extraction methods with mixture models and utilize these features to construct the probabilistic maps and develop an automatic clustering algorithm.