果蠅嗅覺神經影像之三維結構分析及統計分類

Abstract

本研究的主要目的是建構一個自動化流程的三維果蠅嗅覺影像的分類器。此次研究的資料是來自國立清華大學腦科學中心江安世博士所提供的 125張高解析度LSM影像檔，這些影像皆是以果蠅的嗅覺腦區(Antennal Lobe)的方位來命名分成六種類別，分別是DA1、DL1、DL3、DM1、DM2和VL2a。經由扣除可能是實驗染色錯誤的神經影像，剩餘113張影像。由於影像資料有太多雜訊，在此我們藉由對每張神經影像裡的每個物件做標記來達到影像分割的自動化，取出我們需要的神經來達到去雜訊的目的。之後我們針對神經影像取出數種較穩健的特徵值，藉由這些特徵值來區別各種神經影像在空間上的分佈情形。對影像取完特徵值之後，我們對這些特徵值使用逆分層回歸（Sliced inverse regression）可以幫助我們提升分類的正確率。最後使用Weka及R中的SVM,J48,IBk,OneR做統計分類及預測。在此各種分類器的分類結果皆以leave-one-out的cross-validation正確率當做評估的標準。

The goal of our research is to construct an automated process to classify 3D Drosophila calyx images. The 125 high resolution LSM images were administered by Ann-Shyn Chiang from the Department of Life Science at National Tsing Hua University. Those images are classified into six categories that are named by their position in the Antennal Lobe. The six categories are named DA1,DL1,DL3,DM1,DM2 and VL2a. By removing some wrong images that may be caused by experimental errors, there remain 113 images, so we just do a classification on those 113 images. Because the images have too much noise, here we use volume filter to extract useful neurons from images to remove noise automatically. Furthermore, we calculate many robust features based those neuron images. Then we can distinguish different spatial circumstances relative to their dissemination by using those features. After extracting features from images, we use sliced inverse regression on feature data which can help us to increase accuracy. Finally, we use SVM, J48, IBk, and OneR classifiers in Weka and R. Here are different ways to classify results all use leave-one-out cross-validation to evaluate correctness.