- 學位論文
深度學習應用於定量光學顯微術之研究
Deep Learning Applications for Quantitative Optical Microscopy
摘要
傳統光學系統設計須考量硬體限制(如：機械式掃描(Mechanical Scanning))與光學固有物理限制(如：數值孔徑(Numerical Aperture)與繞射極限(Diffraction limit))，導致觀測樣本侷限性高，系統可應用範圍有限，而其主要影響生物研究領域之發展。 近幾十年機器學習(Machine Learning)進展快速，已應用於多個領域，工商農業、交通與醫學均有相關研究幫助發展，而其中深度學習(Deep Learning)更帶來顯著的高效能與高精準度。近5年來，多篇研究應用深度學習於醫學影像處理，提升影像品質或減少必要輻射之劑量，提供除了改善硬體外之有效解決方法。而在光學領域研究中，無透鏡計算成像系統(Lensless Computational Imaging System)、計算相位成像(Computational Phase Imaging)也提供直觀證明，探究深度學習應用於光學系統之可行性。 本論文結合深度神經網路(deep neural network)以及傳統定量光學系統(Quantitative Optical System)，以真實實驗數據影像作為學習資料，訓練並優化神經網路模型參數，達到提升定量光學成像系統之結果，並幫助優化生物研究實驗之觀測方法。 傳統明場(bright-field)光學顯微鏡僅能偵測光線強度分布，因此無法對薄且透明之弱相位物體(weak phase object)(如：細胞、薄組織切片)進行成像，而目前有多種相位對比(phase-contrast)成像術用以重建物體之相位資訊，我們選定定量相位差分對比顯微術(Quantitative Differential Phase Contrast Microscopy, qDPC)作為研究一主軸，旨在優化此定量光學顯微術之成像過程，解決利用非對稱照明(Asymmetric Illumination)蒐集各方向相位資訊時，數位式旋轉光瞳過程所造成無法以高偵率監測活體細胞之限制，在實驗中，深度學習方法僅利用1軸qDPC便能重建出12軸各向同性(Isotropic) qDPC影像，定量相位數值誤差小於5%以內，顯示其有準確定量可行性，並提供10倍以上成像速率。 除此之外，我們選定光場顯微術(Light-field Microscopy)作為另一研究主軸，光場成像不同於傳統，傳統光場成像利用多個相機組成陣列而達到多角度攝影，利用不同角度2D資訊可重建不同聚焦影像，以及取得定量深度3D資訊，可幫助了解生物樣本之表面細微形貌，然而成本考量以及各台相機參數有一定偏差，故發展使用微透鏡陣列(Micro Lens Array)之光場成像方法，微透鏡陣列如同擺放許多針孔於成像面，分開不同角度照明之光線，可實現單次拍攝(Single Shot)便可紀錄光場資訊於2D影像中，然而利用有限像素點數目以紀錄多角度資訊，導致犧牲影像解析度達10倍左右，故於此論文，結合非監督式學習(Unsupervised Learning)之超解析解像術(Super Resolution)與光場顯微術，達到重建不同深度影像資訊，以及維持高影像解析度之結果。 結論，本研究提出以深度學習方法，提升兩種不同定量顯微術之成像結果，一方面為實現高速定量相位成像，另一方面為提升重建不同深度之影像品質。
並列摘要
The performance of the designed optical system is affected by the hardware (i.e. speed and stability of the mechanical scanning) and the inherent optical parameters (i.e. numerical aperture). Typically, the above limitations result in disadvantages for the observation of biological samples for different aspects. Therefore, the scope of application of the system is limited, and it mainly affects the development of the biological research field. In recent decades, machine learning has made rapid progress and has been applied in many research fields. One branch of machine learning called deep learning (DL) has brought significant high performance and accuracy to those researches. In the past 5 years, many studies have applied DL to the medical imaging. In the field of optics, lensless computational imaging systems and computational phase imaging also provide intuitive proofs to explore the feasibility of applying DL to optical systems. This thesis combines deep neural networks with quantitative optical systems. We use the experimental imaging data as learning datasets to train the neural network to achieve the improvement of the quantitative optical imaging systems. The bright-field optical microscope can only detect the light intensity distribution, so it cannot image thin and transparent weak phase objects (such as cells, thin tissue sections). Currently, there are multiple phase imaging methods which are used to reconstruct the phase information of objects. We choose Quantitative Differential Phase Contrast Microscopy (qDPC) as our research topic for its characteristics of scanning-free and high spatial-resolution. The inherent drawback of the qDPC is the requirement of multiple measurements for isotropic phase retrieval. It takes a lot of time to obtain microscope images and compute the quantitative phase. We utilize the DL technique to simplify the imaging procedure of qDPC microscopy and enable the high speed imaging for monitoring different kinds of living cells. In the experiment, the DL model translates the 1-axis anisotropic phase image into the 12-axis isotropic phase image. The results show the predicted phase from our DL model has the error that is less than 5%. The simplified imaging procedure makes the data acquisition speed be 10 times faster than before. Light-Field Microscope (LFM) is one type of microscope for imaging thick biological samples without mechanical scanning. It records light field information in a 2D image with a single shot. 2D sub-views from different angles can be used to reconstruct different depth information. However, it uses a limited number of pixels to record multi-angle information. As a result, the image resolution is sacrificed by N times which relates to the ratio of the diameter of micro-lens and the sensor pixel size. Therefore, in this thesis, the unsupervised super-resolution and the light-field microscopy are combined to reconstruct image information of different depths without losing the contrast. In conclusion, the DL method provides solid evidence for improving the 1-axis qDPC phase uniformity, as well as retrieving missing spatial frequencies. In addition, the unsupervised DL model improves the sub-view image quality for the light-field reconstruction. The proposed DL-based methods in the thesis may help in performing high-resolution quantitative studies for cell biology.