數位液滴聚合酶連鎖反應(Droplet Digital Polymerase Chain Reaction， ddPCR)是一種具高靈敏度與絕對定量能力的核酸檢測技術，可將檢體分割為大量獨立的微反應單元，並根據陽性與陰性液滴的比例，結合柏松分布(Poisson Distribution)統計模型，回推目標核酸濃度。然而，ddPCR的實際應用仍面臨液滴生成穩定性不足、高通量影像分析耗時、液滴定位困難，以及陽性與陰性液滴判讀容易受到人工閾值(Threshold)影響等問題，進而影響濃度估算的準確性與再現性。 本研究建立一套整合微流道液滴生成、雙通道螢光影像、深度學習液滴定位、自動訊號標註、訊號分類與柏松濃度估算的 ddPCR 影像定量流程。參考通道用於確認有效液滴位置與總數，訊號通道則提供對應液滴的螢光特徵。針對大視野影像中液滴尺寸小、數量多且分布密集的特性，建立一套針對高密度小尺寸液滴影像所設計的輕量化深度學習偵測架構，並以分區推論及結果整合完成完整視野分析。訊號分類方面，先以非監督式分群產生候選標註，再利用分群品質指標篩選具清楚群集結構的資料，建立單通道與跨通道訊號分類模型，以降低逐顆人工標註與固定人工閾值的需求。 結果顯示，所提出之輕量化液滴偵測模型的mAP@0.5約為0.990，與原始 YOLOv5s的0.991相近。模型參數量由7.01 M降至0.38 M，浮點運算量由15.8 GFLOPs降至7.8 GFLOPs，權重大小由14.4 MB降至1.7 MB。於20 copies/μL低濃度樣本的初步驗證中，跨通道模型回算濃度為22.64 copies/μL，接近人工參考回算值20.27 copies/μL，並降低直接分群所造成的偽陽性。系統在10⁰–10⁴ copies/μL的實驗條件下呈現良好線性（R² = 0.9851），並完成邊緣運算平台之可行性驗證，此輕量化模型的單次分析能源消耗較原始模型降低約23.1%。本研究所建立之流程提供由液滴生成、影像定位、訊號分類至核酸濃度估算的整合式分析架構，可作為後續自動化與可攜式ddPCR系統發展之基礎。;Droplet digital polymerase chain reaction (ddPCR) is a highly sensitive nucleic acid detection technique capable of absolute quantification. It partitions a sample into a large number of independent microreaction units. It estimates the target nucleic acid concentration from the proportions of positive and negative droplets using a Poisson distribution model. However， the practical application of ddPCR still faces several challenges， including insufficient droplet-generation stability， time-consuming high-throughput image analysis， difficulties in droplet localization， and the susceptibility of positive–negative droplet classification to manually defined thresholds. These limitations may compromise the accuracy and reproducibility of concentration estimation. This study established an integrated ddPCR image-based quantification workflow incorporating microfluidic droplet generation， dual-channel fluorescence imaging， deep-learning-based droplet localization， automated signal labeling， signal classification， and Poisson-based concentration estimation. The reference channel was used to identify the locations and total number of valid droplets， whereas the signal channel provided the corresponding fluorescence features of each droplet. To address the characteristics of large-field-of-view images， in which droplets are small， numerous， and densely distributed， a lightweight deep learning detection architecture specifically designed for high-density， small-droplet images was developed. Full-field analysis was achieved through tiled inference and subsequent result integration. For signal classification， unsupervised clustering was first used to generate candidate labels. Clustering quality metrics were then applied to select datasets with clearly defined cluster structures， which were used to construct single-channel and cross-channel signal classification models. This approach reduced the need for manual droplet-by-droplet annotation and fixed user-defined thresholds. The proposed lightweight droplet detection model achieved an mAP@0.5 of approximately 0.990， comparable to the 0.991 obtained using the original YOLOv5s model. The number of model parameters was reduced from 7.01 million to 0.38 million， computational complexity decreased from 15.8 to 7.8 GFLOPs， and the model weight size was reduced from 14.4 MB to 1.7 MB. In a preliminary validation using a low-concentration sample of 20 copies/μL， the cross-channel model yielded a back-calculated concentration of 22.64 copies/μL， which was close to the manually derived reference value of 20.27 copies/μL， while reducing false-positive classifications caused by direct clustering. The system demonstrated good linearity over the experimental range of 10⁰–10⁴ copies/μL (R² = 0.9851). Its feasibility was also validated on a edge-computing platform， where the lightweight model reduced the energy consumption per analysis by approximately 23.1% compared with the original model. The workflow developed in this study provides an integrated analytical framework spanning droplet generation， image-based droplet localization， signal classification， and nucleic acid concentration estimation. It may serve as a foundation for the future development of automated and portable ddPCR systems.
