提高共軛焦顯微鏡Z軸解析度之多角度影像融合演算法

A Multi-Angle Image Fusion Algorithm for Enhancing the Z-Axis Resolution of Confocal Laser Scanning Microscope

陳冠宇(Kuan-Yu Chen)

指導教授 ： 陳永昌 ； 鐘太郎

國立清華大學/電機資訊學院/電機工程學系/博士(2013年)

摘要

對於立體的研究生物樣本，共軛焦顯微鏡是一種強而有力的工具。相較於其它類型的顯微鏡，共軛焦顯微鏡能提供解析度跟對比都比較好的影像。然而，相較於其橫向的解析度，共軛焦顯微鏡在Z軸方向擁有較差的解析度。這個現象限制了重建之三維生物樣本的體資料在空間上的可信度。為了提高Z軸方向的解析度，其中一種方法為Tilted-view Microscopy。藉由旋轉生物樣本，透過傳統的共軛焦顯微鏡可以獲得從不同視角觀測而得之生物樣本影像。原本因為Z軸解析度較差的關係而無法從單一視角觀察到之生物樣本資訊，現在有機會透過其他視角來取得。這些從不同視角觀測而得之生物樣本影像經過整合即可重建一個擁有相同橫向與縱向解析度之生物樣本體資料。
我們提出了一種影像融合的演算法來整合透過Tilted-view Microscopy所獲得之多角度影像，進而重建出擁有相同橫向與縱向解析度之生物樣本體資料。在這個演算法中，影像堆疊首先經過隨深度變化的反疊積(deconvolution)演算法來修復因為點擴散函數而產生的影像失真。然後，這些經過反疊積處理的影像堆疊再透過以特徵為基礎的對位演算法來進行整合。最後，沒有被這些多角度影像堆疊所擷取之生物樣本資訊則透過以亮度為基礎之內插演算法來進行估測。經過上述步驟以後，我們可以重建一個擁有相同橫向與縱向解析度之生物樣本體資料，其中擁有被多角度影像堆疊所擷取之真實生物樣本資訊和透過估測所產生之原本無法從多角度影像堆疊所獲得之生物樣本資訊。

關鍵字

共軛焦顯微鏡 ； Z軸 ； 解析度 ； 影像融合 ； 反疊積 ； 對位 ； 內插

並列摘要

Confocal laser scanning microscope (CLSM) is a powerful tool for studying biological specimens three-dimensionally. Compared with other microscopes, CLSM can provide images with higher resolution and better contrast. However, the resolution along the Z-axis (the optical axis) is much lower than that along the lateral directions. This phenomenon may hamper the spatial reliability of the reconstructed three dimensional volume data of the specimen. One way to increase the resolution in the Z-axis direction is Tilted-view Microscopy. By rotating the specimen, image stacks from different observation angles can be acquired with conventional CLSM. Missing information that can't be recorded from a single direction, due to the poor Z-axial resolution, can be recorded from other directions. Images derived from different observation angles are then combined to reconstruct one volume data with equal lateral and axial resolutions.
We propose an image fusion algorithm for the multi-angle image stacks derived by Tilted-view Microscopy to reconstruct a 3D volume data of the specimen that has equal lateral and axial resolutions. In this algorithm, image stacks are first deconvoluted with a depth-variant deconvolution method to recover the distortions caused by point spread functions. Then, deconvoluted image stacks are integrated through a feature-based registration algorithm. Finally, an intensity-based interpolation is applied to predict the absent information that is not recorded by these multi-angle images. As a result, a 3D volume data of the specimen with equal lateral and axial resolutions, which has the real points from multi-angle images and the predicted points that are not recorded by these images, is reconstructed.

並列關鍵字

Confocal laser scanning microscope ； Z-axis ； resolution ； image fusion ； deconvolution ； registration ； interpolation

參考文獻

[23] C. C. Wu, G. Y. Chen, Y. C. Chen, H. C. Shao, H. M. Chang, and Y. C. Chen, “Algorithm for the Creation of the Standard Drosophila Brain Model and its Coordinate System,” VIE 2008, pp. 478–483, 2008.

[24] Y. C. Chen, Y. C. Chen, and A. S. Chiang, “Two-level model averaging techniques in Drosophila brain imaging,” Proceeding of 2002 IEEE International Conference on Imaging Processing, vol. 2, pp. 941–944, 2002.

[3] P. Davidovits and M. D. Egger, “Photomicrography of Corneal Endothelial Cells in vivo,” Nature, vol. 244, pp. 366–367, 1973.

[5] A. Egner and S. W. Hell, “Fluorescence microscopy with super-resolved optical sections,” TRENDS in Cell Biology, vol. 15, pp. 207–215, 2005.

[6] S. Herculano-Houzel, “The human brain in numbers: a linearly scaled-up primate brain,” Front. Hum. Neurosci., vol. 3, article 31, pp. 1–11, 2009.