基於YOLO集成模型的巴瑞特氏食道黏膜病變區域偵測研究

巴瑞特氏食道（Barrett’s Esophagus, BE）為食道腺癌（Esophageal Adenocarcinoma, EAC）的主要前驅病變，其早期偵測與診斷對於降低癌變風險至關重要。本研究應用 YOLO（You Only Look Once）模型進行巴瑞特氏食道黏膜病變區域偵測，並進一步開發 YOLO 集成模型以提升偵測效能。研究過程中，我們收集並標註2種 BE影像資料集，針對不同模型架構及參數設定進行訓練與優化，比較各模型在病變區域偵測與分類任務中的表現。本研究旨在發展一套可應用於 BE 影像之病變區域偵測系統，作為臨床預警輔助工具，特別強調召回率（Recall）作為主要效能指標，以反映模型在臨床篩檢中降低漏診風險之實務效益。實驗結果顯示，所提出之 YOLO 集成模型在召回率表現上均優於單一模型，展現其對 BE 病變區域之穩定辨識能力與泛化潛力。本研究證實 YOLO 集成模型在BE影像分析中的效能，可作為輔助診斷工具，減少人工判讀負擔並提升臨床決策效率。未來可進一步將其整合至臨床決策支援系統，以增強 BE 之早期診斷效能。

Barrett’s Esophagus (BE) is the primary precursor lesion of esophageal adenocarcinoma (EAC), and its early detection and diagnosis are crucial for reducing the risk of malignant transformation. This study applies the YOLO (You Only Look Once) model to detect mucosal lesion areas in BE images and further develops a YOLO ensemble model to enhance detection performance. Two types of BE image datasets were collected and annotated, and various model architectures and parameter configurations were trained and optimized. The performance of different models in lesion localization and classification tasks was then compared. The objective of this study is to develop a lesion detection system for BE images that can serve as a clinical early warning support tool, with a particular emphasis on recall as the primary performance metric to reflect the model’s practical utility in reducing missed diagnoses during clinical screening. Experimental results show that the proposed YOLO ensemble model consistently outperforms single YOLO models in terms of recall, demonstrating robust lesion recognition capability and generalization potential for BE detection. The study confirms the effectiveness of the YOLO ensemble model in analyzing BE images, supporting its use as an assistive diagnostic tool to reduce the burden of manual interpretation and improve clinical decision-making efficiency. Future work may focus on integrating the model into clinical decision support systems to further enhance the early diagnosis of BE.