- 學位論文
運用卷積神經網路於顯微鏡影像的人類誘發性多功能幹細胞之自動類別標記之研究
Automatic Class Labeling of Human Induced Pluripotent Stem Cells in Microscopy Images using Convolutional Neural Networks
摘要
本文提出了一種用於顯微鏡圖像中人類誘導多能幹細胞（iPS細胞）的自動類標記系統。該系統使用預訓練的卷積神經網絡（CNN）分類器作為分類的基礎，並產生具有類別概率的彩色編碼圖像。總共有4個類別，每個類別代表人類iPS細胞的編譯增殖過程。在我們的系統中使用的CNN包括卷積層、最大池、平均池化與全連接層，使CNN 架構又被稱為iPSNet。使用25,500個圖像的訓練集和2,400個圖像的測試集來設計和評估。與LeNet和AlexNet相比，我們的結果表明相對較高的準確度（> 95.5％）和較短的執行時間。總之，我們的系統可能被用作幫助生物學家可視化人類iPS細胞增殖過程的工具。
並列摘要
This paper proposes an automatic class labeling system for human induced Pluripotent Stem cells (iPS cells) in microscopy images. The system uses a pre-trained convolutional neural network (CNN) classifier as the basis for classification, and produces color-coded images with class probabilities. There are a total of 4 classes, each of which represents the proliferation process of human iPS cells. The CNN used in our system consists of convolutional layers, max pooling, average pooling, and a fully connected layer, called iPSNet, was designed and evaluated using a training set of 30,000 images and a test set of 2,400 images. Our results demonstrated a relatively high accuracy (>95.5%) and short execution time, when compared with LeNet and AlexNet. In summary, our system could potentially be used as a tool to help biologists visualize the proliferation process of human iPS cells.