基於深度學習的三維空間細胞切割及超像素演算法

3-D Cell Segmentation and Superpixel Algorithms Based on Deep Learning

張捷勝(Chieh-Sheng Chang)

指導教授 ： 丁建均

國立臺灣大學/電機資訊學院/電信工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201902338

摘要

細胞影像切割及超像素切割，皆為影像處理中很重要的課題。前者對於醫學研究及自動化診斷，帶來很大的幫助；後者則廣泛運用在電腦視覺領域中，解決各式各樣的問題。
相較於自然影像的切割，細胞影像的切割更具挑戰性。原因在於影像本身，細胞個體之間的邊界並不明顯，顏色也相似。近年來，有許多以深度學習為基礎的演算法來處理細胞影像切割的問題，像是本篇論文中所使用的V-net架構，便是一個經典的例子。本篇論文中，我們提出一些方法來改善V-net的表現。除了細胞和背景兩個類別，我們額外加入兩種邊界標記，作為不同類別來訓練。由於這些邊界的特性與細胞本體及背景有一定的差異，加入這些標記有助於提升結果。除此之外，我們也採用了基於形態學的後處理方法。藉由以上這些技巧，讓整體的準確率提升，就連影像中對比度較差的區域，也能成功切割出細胞
超像素分割最常被運用在語意切割上。好的切割方法能夠清楚定義不同物體之間的邊界，甚至更進一步能夠切開擁有相似顏色的物體。我們提出一個以最近的一個深度學習模型「超像素採樣網絡」為基礎的方法。我們加入了新的損失函數來加速訓練時的收斂速度。此外，也提出了基於傳統超像素分割演算法的後處理，來提升邊界召回率。我們用實驗顯示了，若與一些以超像素為基礎的語意切割演算法相結合，能達到較高邊界召回率的方法也能夠產生較好的語意切割結果。

關鍵字

細胞影像切割 ； V-net ； 生醫影像處理 ； 超像素切割 ； 語意切割

並列摘要

Cell image segmentation and superpixel segmentation are important topics in image processing. The former benefits for medical research and automatic diagnosis while the latter is widely applied in many other tasks of computer vision.
Cell image segmentation is more challenging than other segmentation problems since cells have similar colors and obscure boundaries. In recent years, the deep learning-based methods, including the V-net, play an important role in image segmentation. In this thesis, we propose several techniques to improve the performance of the V-net for cell segmentation. In addition to cells and background, we add two types of edge labels as different classes to train our network. Since the properties of cell edges are quite different from those of cell bodies and background, these extra labels help improve the performance. Moreover, several morphology-based post-processing algorithms are applied. With these techniques, the accuracy of cell segmentation can be much improved and the cells with poor contrast can still be well segmented.
Superpixel segmentation is mostly applied in semantic segmentation. A good segmentation algorithm can precisely define the boundaries of different objects even with similar colors. We propose a method based on a recent deep learning model, Superpixel Sampling Network. We append a variance loss to the network to speed up the training session. A post-processing method based on a traditional superpixel algorithm is proposed to boost the boundary recall. We also show by experiment that methods with higher boundary recall result in better semantic segmentation results if combined with other superpixel-based segmentation algorithms.

並列關鍵字

cell image segmentation ； V-net ； medical image processing ； superpixel segmentation ； semantic segmentation

參考文獻

[1] S. Becher and C. Lantuéj, “Use of watersheds in contour detection”, in Proceedings of International Workshop on Image Processing, 1979.

[2] S. Beucher and F. Meyer, “The morphological approach to segmentation: the watershed transformation”, in Mathematical Morphology in Image Processing (Ed. E. R. Dougherty), pp. 433–481, 1993.

[3] N. Otsu, "A threshold selection method from gray-level histograms", in IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 9, no. 1, pp. 62-66, 1979.

[4] D. Bradley, G. Roth, "Adaptive Thresholding Using the Integral Image", in Journals of Graphics Tools, vol. 12, no. 2, pp. 13-21, 2007.

[5] J. L. Vincent, "Morphological grayscale reconstruction in image analysis: Applications and efficient algorithms", in IEEE Transactions on Image Processing, vol. 2, pp. 176-201, 1993.