深度學習應用於光學顯微術之研究

Abstract

光學顯微鏡可以通過各種鏡頭和光源的组合來觀察微觀尺度的標本。由於穎體設計和光的物理限制（如繞射極限），系统的性能和穩定性會有所不同，有時可能限制系统能應用的範圍。 基于深度學習的方法被廣泛用於解决逆問题，如相位回推或超分辨率。近期，一種特殊的深度學習演算法:深度圖像先驗被應用於生物醫學影像的計算成像，因為使用這種方法不需要基準真相及大量資料集。深度圖像先驗可以實現自我監督式學習，在此將它予兩種光學系統結合，定量未分相位差顯微鏡和光場顯微鏡，以解决目前技術面臨的問題。 幾乎無法使用傳統亮場顯微鏡觀察薄而透明的待測物，因為這些待測物的相位太小。對於這類待測物，我們使用定量相位差顯微鏡，配合弱相位轉移函數與吉諾洪夫(Tikhonov)正則化，可以自拍攝到的強度影像中計算出待測物的相位分布。雖然目前的方法能有效計算出定量相位，但對於待測物選擇的侷限性和正則化參數調整的不便，影響了該技術的實用性。因此，在論文的第一部分，我們使用深度圖像先驗重建定量相位影像，以解决上述問题。並且使用具有不同相位值的標準樣本，以驗證方法的可行性。 由於透鏡的物理特性，具有高數值孔徑的傳統光學系統的景深通常很短。而光場顯微鏡藉由在相機前插入一個為透鏡陣列来收集更多的角度信息，即使在相同的數值孔徑下也能擁有較長的景深。然而，额外的方向性信息是以空间分辨率為交換獲得的。因此，本論文的第二部分是將深度圖象先驗應用於光場顯微鏡的超级分辨率，以提高空間解析度。空間解析度的提高可以進一步改善影像重建後的品質。實驗中，首先使用解析度測試片，以評估方法的可行性，並將結果與差值方法進行比較。另外，我們使用螢光染色後的三維類器官，以展示系統於生物領域應用的可行性。 結合定量維分相位差顯微鏡和光場顯微鏡中实现的深度圖像先驗可以擴大測物的選擇範圍，免除手動參數調整，並提高空間解析度。自監督學習方法為當前問题提供另一種解决方法，有助於生物醫學成像。

Optical microscopy plays a crucial role in many research areas. Optical microscopes enable the observation of micro-scale specimens by various combinations of lenses and light sources. Due to hardware design and physical limitations of light (such as diffraction limits), system performance and stability would be different and may limit the usefulness of the systems.

The deep learning (DL) based method is extensively used for solving inverse problems such as phase retrieval and super-resolution. Recently, one of the useful networks, deep image prior (DIP), is widely implanted in computational imaging for biomedical images since no exact ground truth or pre-training data are needed. In this thesis, DIP that enables self-supervised learning is adopted in two optical systems, quantitative differential phase contrast (QDPC) microscopy for phase retrieval and light field microscopy (LFM) for super-resolution.

QDPC microscopy with weak phase transfer function (WOTF) and Tikhonov regularization can retrieve the phase distribution of samples from the captured measurements to visualize thin and transparent samples. Although the current method provides an efficient way for phase retrieval, the selection of thin samples and the tedium in regularization parameters tuning limit the usefulness. Therefore, in the first part of the thesis, we propose a DIP algorithm adapted to QDPC microscopy for phase retrieval to solve the above-mentioned issue. To verify the feasibility of the proposed method, standard phase targets with various phase values are used.

Due to the physical properties of the lens, the depth of field (DoF) of conventional optical systems with a high numerical aperture (NA) is usually short. LF microscopy, on the other hand, has a larger DoF even with the same NA. By simply inserting a micro-lens array in front of the CCD, the imaging system can collect extra angular information to expand the DoF. However, additional directional information is obtained at the cost of spatial resolution. Therefore, to enhance the spatial resolution LFM, we proposed a DIP algorithm in the second part of the thesis. The resolution target is first used for evaluating the proposed method, and fluorescent organoids are used to show the ability for biological application.

DIP implemented in QDPC microscopy can extend sample selection and eliminate manual parameter tuning. The implementation of the LF microscope can improve spatial resolution. The self-supervised learning method provides an alternative way of solving optical problems.