本研究旨在建立光譜分析與生醫電腦視覺技術之人工智慧（Artificial Intelligence， AI）辨識流程，以解決臨床上在大量周邊血白血球（White Blood Cells， WBCs）背景中快速初篩稀有循環腫瘤細胞（Circulating Tumor Cells， CTCs）之核心痛點。傳統CTC檢測方法普遍面臨繁瑣標記、高成本與時效性不足等限制。因此本研究以維持細胞活性之無染色（unstained）檢測為核心，建構了包含光譜分類、靜態影像分類與動態細胞追蹤之三種分析流程： 首先，在光譜分析中，針對肺腺癌細胞株（A549與H1299）與白血球高維雜訊及基線漂移之挑戰，建立了一套依序導入多項式擬合（Polynomial Baseline Correction）與經驗模態分解（Empirical Mode Decomposition， EMD）之基線校正、中值濾波降噪（Median Filtering）、多幀非局部均值降噪（Multi-frame Non-Local Means， MNLM）等前處理流程，並透過隨機森林（Random Forest， RF）、極端梯度提升（Extreme Gradient Boosting， XGB）與支援向量機（Support Vector Machine， SVM）等演算法與是否加入主成分分析（Principal Component Analysis， PCA）進行交叉驗證，在MNLM + PCA + SVM 的組合下，可於 A549/WBC 及 H1299/WBC 資料集分別達到0.878 ± 0.054與0.928± 0.061的準確率，並藉由PCA Loading 分析提升模型之可解釋性，降低黑箱效應。 其次，在靜態生醫影像辨識方面，利用倒立式顯微鏡取得 40× 無染色相位差影像，並導入具備像素級語義分割能力之 Mask R-CNN（Mask Region-based Convolutional Neural Network） 深度學習模型。為建立客觀之真實標籤（Ground Truth），實驗中引入雙光路螢光金標準驗證機制。實驗結果顯示，該模型能有效克服細胞局部聚集與邊界重疊現象，在不同的細胞混合比例下，A549 腫瘤細胞的辨識率分別達到 88.23%（1:1）、96.77%（1:5）與 100%（5:1）。 最後，在動態細胞追蹤方面，利用注射幫浦控制微流道主通道內之流體環境，以 30 fps 擷取細胞流經影像。為符合臨床即時篩檢之需求，本研究結合限制對比度自適應直方圖均衡（Contrast Limited Adaptive Histogram Equalization， CLAHE）、霍夫圓轉換（Hough Circle Transform， HCT）與 Trackpy 細胞追蹤演算法，達到細胞定位與運動軌跡追蹤之成果。 綜上所述，本研究針對光譜、靜態影像與動態影像三類生物資訊，分別建立對應之人工智慧分析流程，以實現不同資料型態之辨識。其中，光譜分析使用機器學習模型根據光譜訊號進行細胞分類；靜態影像分析流程結合細胞分割與分類技術，提升細胞辨識能力；動態影像分析流程則透過物件追蹤技術，完成微流道中細胞位置之追蹤。上述三種分析流程皆以無染色、低成本之資料擷取方式為基礎，並藉由人工智慧技術提升細胞分析之自動化程度與效率，增加其應用於循環腫瘤細胞快速篩檢之潛力。;This study aims to establish artificial intelligence (AI)-based analytical workflows incorporating spectral analysis and biomedical computer vision to address the clinical challenge of rapidly prescreening rare circulating tumor cells (CTCs) from a large background of peripheral white blood cells (WBCs). Conventional CTC detection methods are generally limited by labor-intensive labeling procedures， high costs， and insufficient detection efficiency. Therefore， an unstained detection strategy that preserves cell viability was adopted as the core principle， and three analytical workflows， including spectral classification， static image analysis， and dynamic cell tracking， were established. For spectral analysis， a preprocessing pipeline was developed to address the challenges of high-dimensional noise and baseline drift in Raman spectra acquired from the lung adenocarcinoma cell lines (A549 and H1299) and WBCs. The proposed pipeline sequentially applied baseline correction using polynomial fitting and empirical mode decomposition (EMD)， followed by median filtering and multi-frame non-local means (MNLM) denoising. Subsequently， Random Forest (RF)， Extreme Gradient Boosting (XGB)， and Support Vector Machine (SVM) classifiers were evaluated， both with and without principal component analysis (PCA)， using cross-validation. Among the evaluated workflows， the combination of MNLM + PCA + SVM achieved the best classification performance， yielding classification accuracies of 0.878 ± 0.054 and 0.928 ± 0.061 for the A549/WBC and H1299/WBC datasets， respectively. Furthermore， PCA loading analysis was performed to investigate the spectral features， thereby improving model interpretability and reducing the black-box nature of the classification model. For static biomedical image analysis， unstained phase-contrast images were acquired using an inverted microscope at 40× magnification， and a Mask Region-based Convolutional Neural Network (Mask R-CNN) was employed to perform pixel-level instance segmentation. To establish objective and reliable ground-truth annotations， a dual-optical-path fluorescence validation strategy was introduced as the reference standard. The experimental results demonstrated that the proposed model effectively addressed local cell aggregation and overlapping cell boundaries commonly observed in adherent cells. Under different WBC-to-A549 mixing ratios， the identification rates of A549 tumor cells reached 88.23% (1:1)， 96.77% (1:5)， and 100% (5:1). For dynamic cell tracking， a syringe pump was used to control the fluid environment within the main channel of the microfluidic device， and cell flow images were acquired at 30 fps. To meet the requirements of real-time clinical screening， the proposed workflow integrated Contrast Limited Adaptive Histogram Equalization (CLAHE)， the Hough Circle Transform (HCT)， and the Trackpy cell-tracking algorithm to achieve cell localization and trajectory tracking. In summary， three AI-based analytical workflows were established for the analysis of Raman spectra， static microscopic images， and dynamic cell migration images， respectively. In the spectral workflow， machine learning models were employed to classify cells based on Raman spectral signatures. In the static image workflow， cell segmentation and classification techniques were integrated to improve cell identification performance. In the dynamic image workflow， object tracking techniques were applied to continuously track cell positions within microfluidic channels. All three workflows were developed using unstained， low-cost data acquisition strategies and incorporated artificial intelligence to improve the automation and efficiency of cell analysis， thereby demonstrating their potential for the rapid screening of circulating tumor cells (CTCs).
