果蠅觸角葉中嗅小球的螢光顯微影像分割，對於研究與建立果蠅腦部連結體（connectome）十分重要。對此，先前的研究曾發展多種方法，並應用於果蠅（Drosophila melanogaster）腦部影像之研究，而大多數的方法都需要建立果蠅標準腦（standard brain）並導入各種相關影像。然而，在這樣的流程中，涉及影像對位（image registration）的技術，而影像對位需要扭轉或移動影像，並有可能造成神經位置的失真；為了繞過這些問題，並有利果蠅腦部連結體之研究，我們發展了一套自動化工具，進行影像分割，並可在果蠅腦部螢光顯微影像中分割出一些區域，而不需要在標準腦中進行影像對位。
在我們的方法中，採用了包含 2 階段 Unet 訓練的方法。我們首先訓練了第一個 Unet 的模型，並以此初步劃分腦區的輪廓；之後訓練了第二個 Unet，用以填補第一個 Unet 畫出的輪廓所產生的破洞，並輔以其他方法如區域成長演算法（Region Growing method），分割出腦區的邊界輪廓。在完成二維影像分割後，我們以這些二維的腦影像重新建構三維腦區；我們計算連續切片中，相鄰切片的每個嗅小球的質心之間的相對距離，作為三維腦區影像的重建。經過測試與分析，在 49 顆嗅小球中，我們成功的建構出了 28 顆嗅小球。
本研究中的流程經自動化，未來將可用於果蠅全腦的顯微螢光影像之自動化分割。

Identifying the innervated brain regions of each neuron is crucial for constructing the connectome. In previous studies, several approaches have been developed and applied to the brain images of fruit fly (Drosophila melanogaster), and most of the approaches require establishing a standard brain and registering every image into it. However, image registration involves shifting and warping, and may cause substantial dislocation of a neuron from its true location. In order to address this issue, we developed a method for automated brain image segmentation and used this method to identify brain regions in images without registering them to the standard brain.
In the proposed method, we adopted a two-stage Unet approach. We first trained a Unet model to recognize the boundaries of brain regions and then trained a second Unet model to fill the holes on the boundaries left from the first Unet. Next, we implemented other algorithms, e.g. Region Growing methods, to segment the brain regions from their boundaries. After the completion of 2D image segmentation, we reconstructed the 3D brain regions from their 2D stacks based on the proximity of the regions in consecutive slices. We tested the proposed method in segmenting glomeruli in Antennal Lobes of fruit flies and successfully reconstructed 28 out of 30 glomeruli. The pipeline built in this study could be further developed for the segmentation of the fluorescent images of the whole fly brain in the future.

1. de Bruyne, M., K. Foster, and J.R. Carlson, Odor Coding in the Drosophila Antenna. Neuron, 2001. 30(2): pp. 537-552.
2. Hallem, E.A. and J.R. Carlson, The odor coding system of Drosophila. Trends in Genetics, 2004. 20(9): pp. 453-459.
3. Huang H. et al., Unet 3+: A full-scale connected Unet for medical image segmentation. IEEE ICASSP, 2020. 5(20): pp. 1055-1059.
4. Kaur, D. and Y. Kaur, Various Image Segmentation Techniques: A Review. IJCSMC, 2014. 3(5): pp. 809-814.
5. Keene, A.C. and S. Waddell, Drosophila olfactory memory: single genes to complex neural circuits. Nat Rev Neurosci, 2007. 8(5): pp. 341-354.
6. Kirilliv A. et al., PointRend: Image Segmentation as Rendering. IEEE CVPR, 2020. pp. 9799-9808.
7. Laissue, P.P., et al., Three-dimensional reconstruction of the antennal lobe in Drosophila melanogaster. Journal of Comparative Neurology, 1999. 405(4): pp. 543-552.
8. LeCun, Y., Y. Bengio, and G. Hinton, Deep learning. Nature, 2015. 521: p. 436.
9. Li, J.-B., et al., An algorithm for Drosophila glomerulus image segmentation. 2019 (unpublished).
10. Lin, H.-H., et al., A Map of Olfactory Representation in the Drosophila Mushroom Body. Cell, 2007. 128(6): pp. 1205-1217.
11. Masse, N.Y., G.C. Turner, and G.S.X.E. Jefferis, Olfactory Information Processing in Drosophila. Current Biology, 2009. 19(16): pp. R700-R713.
12. Minaee S., et al., Image Segmentation Using Deep Learning: A Survey. Preprint arXiv:2001.05566v5, 2020.
13. Ronneberger, O., P. Fischer, and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation. Springer MICCAI, 2015. pp. 234-241.
14. Soille, P. and L. M. Vincent, Determining watersheds in digital pictures via flooding simulations. Visual Communications and Image Processing '90, 1990. 1360(5): pp. 240-250.
15. Zhou Z. et al., UNet++: A Nested U-Net Architecture for Medical Image Segmentation. Preprint arXiv:1807.10165v1, 2018.