| 摘要: | 本研究旨在發展新穎的高通量顯微術(High-throughput microscopy, HTM)，結合計算光學(Computational Optics)與深度學習(Deep Learning, DL)兩項技術應用在細胞影像等相關研究。高通量顯微技術具有可擷取大面積並保有高解析度影像之特性，因此可同時取得多種生理條件與環境下之巨量的細胞結構以及其行為動力學的時間與空間資訊[1-3]，經過減少個體差異性及可系統化的生物過程中進行計算評估[4,5]，可取得傳統的生物顯微術無法量測與估算的大量或群體細胞生物過程之隱藏特徵。透過本研究預計發展的深度學習顯微術(Deep Learning Microscopy, DLM)，可以解決當前HTM面臨的技術挑戰。研究技術核心為使用傅立葉疊層成像顯微術(Fourier Ptychographic Microscopy, FPM)[10-12]，調變不同的照明角度以產生具有不同空間移位的光學信息，透過收集低分辨率圖像序列後，重建取得空間高頻分量之資訊，進而同時取得大視場與高分辨率影像。在我們初步實驗中，已經成功重建一組測試樣本的大視場(Field of View,FOV)與高分辨影像。另外，本研究計畫結合深度學習(DL)技術發展一套深度學習顯微系統(DLM)，可透過深度學習在大型數據集中獲得與分析信息的能力[13,14]，應用於高通量影像分析，進行進行癌細胞之篩選抗癌藥物和鑑定耐藥機制之研究。發展深度學習顯微術(DLM)提供相關癌症研究新面向技術突破，尤其是在篩選抗癌藥物和鑑定耐藥機制方面，此技術不僅能同時處理的大視場(FOV)與高分辨率的影像，也大幅降低傳統顯微技術所需的系統複雜度和成本，使研究者在一般實驗室環境下仍可使用[15]。 ;High-throughput microscopy (HTM) is an indispensable tool for cancerous research. However, conventional microscopes can achieve either a large field of-view (FOV) or high resolution, but not both simultaneously. Thus, this proposal aims to develop a new HTM approach, aiming to integrate the merits of computational optics and deep learning microscopy. Visualizing a large number of individual cells, HTM generates vast spatiotemporal information on cellular structures and behavioral dynamics under a multitude of physiological condition[1-3]. It computationally evaluates the whole aspect of such data, yielding unbiased, systematic representation of biological process[4,5], and hidden features in bulk or small population measurements. This proposal implements computational illumination approach, known as Fourier ptychographic microscopy (FPM)[10-12]. FPM collects low resolution image sequences and changes the position of a point-light source; each image has spatially-shifted spectrum information in Fourier space, where the amount of shift depends on the illumination angle. By numerically stitching image sequences, the Fourier space, including high frequency components, can be reconstructed, which enables FPM to achieve high spatial resolution with low magnification lenses. In our year 1 pilot study, we successful constructed an FPM system to detect sample pattern. The promising result shows both large field-of-view and high resolution. In addition, this proposal adopts deep learning technology to develop deep learning microscopy (DLM). Deep learning (DL) is making an impressive progress in its ability to derive information from large data sets, to the point where such techniques can outperform human analyses for many data sets[13,14]. DL is particularly potent in discovering intricate, hidden structure in high dimensional data sets, and has been adopted for solving quantum physics problem. This proposal will leverage DL power to analyze large images from FPM. We will develop a DLM platform with the capacities for i) high-resolution, wide FOV imaging; ii) molecular profiling on individual cells; and iii) automated imaging analyses (molecular expression, morphology, migration, propagation) on large numbers of single cells. The development of DLM will have unprecedented analytical power: imaging large number of individual cells at high spatial resolution, and automatically extracting multitude of cellular features. As such, we can envision that the DLM will be a transformative tool for cancer research. Potential applications include better monitoring anticancer drug responses, analyzing cellular heterogeneity in large section of tissues, and prospectively detecting cellular fate under various physiological perturbations[15]. |
