三維果蠅嗅覺神經影像之統計分類

Abstract

本研究主要目的是建構果蠅嗅覺神經影像的分類器，在此我們有六種三維影像都以果蠅的嗅覺腦區 (Antennal Lobe)來命名，分別是DL1, VL2a, DM1, DM2,DL3 和DA1。一般而言，影像資料有太多冗贅訊息，我們針對每一個神經路徑萃取不同的特徵用以描述該三維圖像中的神經路徑複雜程度，並以這些特徵來發展分類器。在此，發展的分類器皆以Leave-one-out的cross-validation 正確率當作評估的標準。在本研究中，分成六類的分類器中最好的正確率是54.4%，比亂猜的正確率1/6高出三倍多，在此加入旋轉骨架的端點數特徵並將影像強度值的相對頻率取對數會幫助提升正確率。

The main study purpose is the application of a classification method for six kinds of 3D Drosophila (fly) Calyx Images automatically. We have six different classes, which are named by the glomeruli in Antennal Lobe, DL1, VL2a, DM2, DM1, DL3 and DA1. Generally speaking, most of the image data contain redundant information; the extracted features describing the 3D olfactory neuron pathway of the six calyxes will help us construct a classification method. On the other hand, the classification cross validation accuracy is helpful to determine the essentialness of these features. First, SVM classifiers outperform better than LDA classifiers across the 23 different features combinations in accuracy. Secondly, the 18th model (six-category-SVM) has the highest leave-one out cross validation accuracy, 54.4% (more than three times of random guess). Rotational skeleton Endpoint feature helps the six-category- SVM classifier with log relative frequency vector and histogram feature in the 18th model.