| 摘要: | 本研究基於數位同軸全像術（Digital In-line Holography），建構一套具備多波長、廣視野與深度學習重建能力之無透鏡顯微影像系統，旨在突破傳統光學顯微鏡於體積、成本與視野限制上的瓶頸。系統採用紅光（650 nm）、綠光（532 nm）與藍光（406 nm）三波長LED光源，經由光纖耦合匯聚至單一出射點，搭配高解析度CMOS感測器，形成近同軸光學架構。此設計有效簡化鏡頭與對準機構，單次曝光可擷取約30 mm²之繞射影像，視野面積約為傳統20倍物鏡的186倍，展現其於大面積樣本觀測與廣視野成像上的優勢。
針對同軸全像架構常見之 twin image 干擾與相位重建穩定性不足問題，本研究提出雙距離影像擷取策略，於樣品與感測器間設置兩組記錄距離，並結合雙向傳播與物理一致性約束演算法，以提升相位重建的準確性與解析度。進一步導入深度學習模型，取代傳統數學傳播過程，以高速重建振幅與相位影像，並整合去雜訊與細節增強策略，提升影像品質與跨樣本穩定性。
系統針對三波長影像進行獨立重建後再進行融合，重建擬真彩色影像，無需染色濾光片或鏡頭，即可呈現樣本於不同光譜下的形貌特徵與色彩對比。實驗結果顯示，本系統能準確還原染色組織切片中之絨毛結構、黏膜層次與細胞核排列，與實際顯微觀察結果高度一致。針對大尺寸影像處理需求，亦開發重疊裁切與無縫拼接演算法，搭配固定輸入尺寸之深度學習模型，有效完成整體影像的連續重建與全視場拼接。
綜上所述，本系統兼具高解析、廣視野與彩色擬真之顯微成像能力，未來可望應用於組織病理分析、細胞分類、影像診斷輔助與生醫顯微成像等多元場域，為數位病理與遠距醫學診斷提供具可攜性與模組化潛力之創新成像平台。
;This study presents the development of a lensless microscopy system based on Digital In-line Holography (DIH), featuring multi-wavelength illumination, wide field of view, and deep learning–assisted image reconstruction. The system is designed to overcome the limitations of conventional optical microscopes in terms of size, cost, and field coverage. It employs three discrete wavelength LED sources—red (650 nm), green (532 nm), and blue (406 nm)—which are coupled via a 3-to-1 optical fiber into a single emission point. A high-resolution CMOS sensor is used in near-coaxial configuration, simplifying lens alignment and optical complexity. This configuration enables the acquisition of diffraction patterns over an area of approximately 30 mm² in a single exposure, equivalent to 186 times the field of view of a conventional 20× objective lens, demonstrating its advantages for wide-area and high-throughput imaging.
To address the typical challenges of DIH, including twin image artifacts and unstable phase retrieval, this work introduces a dual-distance acquisition strategy by capturing holograms at two different axial positions between the sample and sensor. Combined with a bidirectional propagation algorithm and physical-consistency constraints, the approach enhances phase reconstruction accuracy and spatial resolution. A deep learning model is further incorporated to replace traditional numerical propagation, enabling rapid amplitude and phase reconstruction while integrating denoising and detail enhancement strategies to improve image quality and cross-sample stability.
Each wavelength channel is reconstructed independently and then fused to form a without the use of optical filters or staining lenses. The system accurately restores structural features of stained tissue slices—such as villi, mucosal layers, and nuclear distributions—with high consistency to standard microscopy observations. To handle large-scale raw data, a patch-based stitching algorithm is developed to enable seamless global image reconstruction with fixed-size model inputs.
In summary, the proposed system achieves high-resolution, wide-field, and color-representative microscopic imaging and shows promising potential for applications in histological analysis, cellular classification, diagnostic imaging platforms, and digital pathology. |
