利用Dirichlet混合模型自動化建立彩虹腦神經機率圖

Abstract

彩虹腦(Brainbow)技術因為可以使神經細胞帶有不同的色彩而被科學家用來試圖辨認大腦的神經網絡。然而，彩虹腦影像有著全面性的混疊現象(crosstalk effect)需要被解決進而得到正確的神經影像。此外，由於神經系統交錯複雜，我們需要建立自動化擷取神經影像的技術來節省研究者的時間。在這篇研究當中，我們利用Dirichlet混合模型來對影像的顏色資訊進行分析，並自動產生擷取神經影像的結果。而我們也對Dirichlet混合模型的參數估計提出了一個改進的方法來求得更佳的估計值。最後，由於影像品質的限制，有些原本應該相連的神經影像會斷開。我們提出區塊成長法(region growing)來使這些神經影像有機會再連結在一起。結合這些方法，我們建構出一套自動化的分析方法來擷取彩虹腦中的神經。這將可以幫助科學家更容易從原始的彩虹腦影像中發現更有趣的資訊。

The Brainbow technology proposed to help scientists identify the brain circuit by making the brain neurons have color expressions. However, the Brainbow images suffer from a crosstalk effect that is needed to resolve. In addition, to save researchers’ time, an automatic extraction method should be developed to find the neuron components from the interknitted nervous system. In this study, we solve these issues by fitting the information of the Brainbow image a Dirichlet mixture model. An improved ECM method for the Dirichlet mixture model is also developed to give better estimates. Finally, a region growing approach is proposed to find the neuron components from some disconnected components that due to restricted quality of the images. With this systematic methodology, researchers can find more interesting results in the Brainbow image more easily.