結合人工智輔助之定量微分相位差顯微術於虛擬組織學染色之研究

Abstract

各向同性定量微分相位差顯微術( Isotropic qDPC ) 是一種免標記定量相位影像技術，可利用非對稱照明將薄且透明樣本的相位梯度資訊轉換為可量測的強度變化，再透過弱物體轉移函數（weak object transfer function, WOTF）與相位重建演算法，如吉諾洪夫（Tikhonov）正則化方法重建樣本的高解析度定量相位影像，適用於長時間的活細胞動態檢測及臨床病理與生命科學相關研究。本論文目標為運用Isotropic qDPC系統在組織學染色之成像，分別從顯微鏡系統模組、深度學習虛擬組織學染色方法及生物醫學應用三個層面進行討論。 在本研究中，結合光學系統、自動軟體介面控制以及後處理影像模組，執行三個基本功能：照明控制、影像採集和影像重建。利用自動化 LabVIEW介面整合控制薄膜電晶體液晶顯示器（TFT panel）即時輸出設計之照明圖樣，與 MATLAB 相位重建模組，以實現可程式化照明圖樣控制、即時影像擷取與相位重建流程。Isotropic qDPC 顯微系統驗證結果顯示，本系統可用於標準相位樣本、細胞樣本與病理切片之免標記相位成像，並可提供後續深度學習分析所需的相位影像輸入。 然而，傳統 qDPC 相位重建多仰賴 Tikhonov 正則化，其正則化參數會影響相位振幅、背景平滑程度與細微結構保留，進而限制後續病理虛擬染色模型對不同相位重建條件的穩定性。因此，本論文進一步提出自監督式 qDPC 相位重建模型，結合深度影像先驗（DIP）、注意力 U-Net 神經先驗與可微分 qDPC 物理前向模型，以降低針對不同樣本需人為調整 Tikhonov 正則化參數的依賴。此相位重建結果進一步作為下游虛擬組織學染色模型的輸入，以評估物理限制相位重建是否可提升虛擬染色表現。 在臨床病理與生命科學研究中，組織學染色（histological staining）為觀察組織形態、細胞結構與病理特徵的標準方法。然而，傳統化學染色需經固定、包埋、切片、去蠟、染色與封片等多步驟染色與樣本處理，且化學染色具有破壞性，染色過程限制同一切片後續進行額外染色、分子檢測與基因分析。因此，本研究發展深度學習虛擬組織學染色（virtual histological staining）方法，使用生成式對抗網路（generative adversarial network, GAN）將未染色、無標記（label-free）的各向同性定量微分相位差顯微影像轉換為具有組織病理染色對比的影像。此方法可在保留樣本原始形態資訊的同時重建具有組織學對比的影像，提供虛擬多重染色（virtual multiplexed staining）。對於腎臟腫瘤或疑似癌細胞分析，虛擬染色可先用於大範圍篩選細胞密集區域，後續可以針對特定區域進行後續基因檢測與分子分析，並降低樣本處理所需時間與成本。 生物醫學應用從影像形態、組織結構特徵資訊及深度學習模型表現分析等方面探討可結合之人工智慧模型於疾病診斷與生物醫學相關應用，初步研究以糖尿病腎病變（diabetic nephropathy, DN）、抗中性球細胞質抗體相關血管炎（anti-neutrophil cytoplasmic antibody-associated vasculitis, ANCA-associated vasculitis, AAV）與微小變化病（minimal change disease, MCD）的腎臟病理切片觀測。首先使用傳統 Tikhonov-qDPC 影像建立循環一致性生成對抗網路（CycleGAN）、對比式非配對影像轉換（CUT）與可解釋式對比非配對影像轉換（E-CUT）等無配對影像轉換基準模型。結果顯示，E-CUT 在人類腎臟相鄰切片資料集中具有較佳的整體虛擬 H&E 與 PAS 染色表現，因此被選為後續模型改良的基準架構。 最後，本研究將自監督 qDPC 相位重建與病理基礎模型引導之表徵轉移結合，提出 Pathology-constrained Self-supervised qDPC E-CUT 架構。此方法將 RetCCL 預訓練病理編碼器整合至 E-CUT 生成器中，使模型能利用大規模病理切片資料所學得的多尺度組織學特徵。研究結果顯示，與傳統的亮場顯微鏡相比，定量微分相位差成像可獲得組織外觀型態以及內部結構等資訊用以評估細胞條件。而相較於原始 E-CUT 與僅使用自監督 qDPC 輸入的 E-CUT，Pathology-constrained E-CUT 在 FID 與 KID 指標上取得最佳整體表現，並在不同 qDPC 相位重建條件下展現較佳的穩定性。本研究顯示，結合 qDPC 免標記相位影像、自監督式物理限制相位重建與病理表徵轉移學習，展現其應用於人工智能輔助生物醫學影像與數位病理之潛力。

Quantitative phase imaging (QPI) has become an important tool in biomedical research for observing thin and transparent biological specimens. However, obtaining accurate structural information from such samples remains challenging using conventional bright-field microscopy due to the lack of intrinsic contrast. Isotropic Quantitative Differential Phase Contrast (Isotropic qDPC) Microscopy is a label-free quantitative phase imaging technique that converts phase-gradient information of transparent samples into measurable intensity variations through asymmetric illumination. By employing the weak object transfer function (WOTF) and Tikhonov regularization, high-resolution quantitative phase images can be reconstructed, enabling long-term live-cell observation and applications in life science and clinical pathology. In this thesis, an isotropic qDPC microscopy system was developed based on a commercial inverted microscope, a programmable thin-film transistor (TFT) panel, an automated LabVIEW control interface, and a MATLAB-based phase reconstruction module. The system integrates illumination pattern control, image acquisition, and quantitative phase reconstruction, enabling label-free phase imaging of standard phase targets, biological cells, and renal pathology sections. A major limitation of conventional qDPC phase retrieval is its dependence on manually selected Tikhonov regularization parameters across different specimens. Different regularization settings can alter phase amplitude, background smoothness, and fine structural details, which may further affect downstream virtual histological staining. To address this limitation, this thesis developed a self-supervised qDPC phase-retrieval framework based on Deep Image Prior, an Attention U-Net neural prior, and a differentiable qDPC physical forward model. This model was designed to reduce the dependence on hand-tuned Tikhonov regularization and to generate more stable pathology-oriented qDPC phase maps for downstream virtual staining. For virtual histological staining, this study focused on renal pathology specimens, including diabetic nephropathy (DN), anti-neutrophil cytoplasmic antibody-associated vasculitis (ANCA-associated vasculitis, AAV), and minimal change disease (MCD). Conventional Tikhonov-reconstructed qDPC images were first used to establish baseline unpaired image-translation models, including Cycle-Consistent Generative Adversarial Network (CycleGAN), Contrastive Unpaired Translation (CUT), and Explainable Contrastive Unpaired Translation (E-CUT). Among these models, E-CUT showed the strongest overall baseline performance for virtual H&E and PAS staining in the human renal adjacent-section dataset. Building on this baseline, a Pathology-constrained Self-supervised qDPC E-CUT framework was further developed by combining self-supervised qDPC phase reconstruction with pathology foundation-model-guided representation transfer. A RetCCL-pretrained histopathology encoder was integrated into the E-CUT generator to provide multiscale pathology-aware features learned from large-scale whole-slide image pretraining. The final comparison showed that Pathology-constrained E-CUT achieved the best overall FID and KID performance among the evaluated variants and demonstrated improved robustness under different qDPC reconstruction conditions. These results suggest that qDPC-based virtual histological staining benefits jointly from stable optical phase reconstruction and pathology-aware representation learning. The proposed framework shows potential for AI-assisted label-free digital pathology and renal virtual histological staining.