基於深度學習的Hep-2細胞影像自動分類技術在自體免疫疾病診斷中的應用

Application of Deep Learning-Based Automatic Classification of Hep-2 Cell Images in Autoimmune Disease Diagnosis

抗核抗體檢驗（anti-nuclear antibody test, ANA test）是診斷自體免疫疾病的重要工具，在辨識抗核抗體型態（ANA pattern）的部分，臨床上常有混合抗核抗體型態（ANA mixed pattern）的案例，也就是說被螢光染色的Hep-2細胞不只一種型態，讓辨識型態的工作增加了複雜性，且形態判斷上容易因經驗不足、主觀意識，而提高報告不一致性。本研究首要目標為利用公開的預訓練卷積神經網絡（CNN）模型建構能辨識混合抗核抗體型態（ANA mixed pattern）的模型，且表現能與資深醫檢師的辨識結果高度一致。次要目標是探討以常見混合型態和單一型態作為模型分辨種類，與使用多標籤分類法相比，本研究所提的訓練模式是否可以提升影像辨識表現。
本研究收集了2020年1月至2023年5月雙和醫院醫學檢驗科的1894張ANA test圖檔，整理出包含單一型態與混合型態共11類。利用深度學習的預訓練CNN模型InceptionResNetV2，建構能辨識ANA混合型態的模型，並與一位資深醫檢師和兩位初學者進行ANA型態辨識的一致性評估，衡量模型表現的兩個重要指標為：Mean Class Accuracy（MCA）和Kappa係數（Cohen's Kappa）。MCA用於評估分類模型的整體表現，它通過對每個類別的分類準確率進行平均，避免因某些類別樣本數較多而對整體準確率產生過大的影響。Kappa係數則是一種統計指標，用於評估兩個模型在分類任務中的一致性，特別是在本研究中，它用來衡量所提出的模型與人員之間的一致性。
研究結果顯示，所建置的模型mean class accuracy（MCA）最高達87.8%。在混合型態的分類效能較差，可能原因有兩個：一是混合型態資料量相對較少，導致模型學習成效不佳；二是混合型態包含兩項特徵，若特徵表現偏頗於其中一項，可能導致誤判。AI模型與資深醫檢師的一致性Kappa係數達85.5%，顯示不錯的一致性，且模型表現優於初學者（83.6% vs 65.5%）。
利用深度學習技術建構的模型能有效輔助醫檢師進行ANA型態辨識，尤其在資源有限的情況下，可提升臨床診斷的準確性和效率。

The anti-nuclear antibody test (ANA test) is vital for diagnosing autoimmune diseases. However, identifying mixed ANA patterns (ANA mixed patterns), where Hep-2 cells display multiple patterns, complicates the process and increases reporting inconsistencies due to subjective judgment or inexperience. This study aims to develop a model using pre-trained convolutional neural networks (CNN) to accurately recognize ANA mixed patterns, with performance comparable to experienced technologists. It also explores whether focusing on common mixed and single patterns improves recognition compared to multi-label classification. We collected 1,894 ANA test images from Shuang Ho Hospital (January 2020 to May 2023), classifying them into 11 categories. Using the pre-trained InceptionResNetV2 model, we built an AI model to identify ANA mixed patterns, evaluating its consistency against one senior and two novice technologists. Performance was measured using Mean Class Accuracy (MCA) and Cohen’s Kappa. MCA averages classification accuracy across classes, preventing overrepresentation by large-sample classes, while Cohen’s Kappa assesses agreement between models. The results show that the constructed model achieved a maximum Mean Class Accuracy (MCA) of 87.8%. However, the classification performance for mixed patterns was relatively weaker, likely due to two reasons: (1) the smaller amount of mixed-pattern data, which affected the model's learning efficiency; and (2) the inclusion of two features in mixed patterns, where the model might lean toward one feature, leading to misclassification. The consistency between the AI model and the senior medical technologist, as measured by the Kappa coefficient, was 85.5%, indicating a high level of agreement. Moreover, the AI model outperformed the novice technologists (83.6% vs. 65.5%). In conclusion, the AI model built using deep learning techniques can effectively assist medical technologists in identifying ANA patterns, especially in resource-limited settings, thereby improving the accuracy and efficiency of clinical diagnosis.