具備年齡動態閾值調控之多模態卵巢癌輔助診斷網頁系統：結合影像蒸餾與異質集成學習

Abstract

背景：卵巢癌因發病隱匿且缺乏早期特異性症狀，導致致死率偏高。傳統診斷高度仰賴 CA-125 或超音波等單一模態，常面臨特異度受限與主觀偏差之侷限。此外，傳統系統多依賴單一且固定的診斷閾值，在停經前與停經後之族群間易產生年齡干擾偏差。 目的：本研究旨在設計、實作並評估一套多模態臨床決策支援系統（CDSS），應用於卵巢癌之術前診斷，以期提升不同年齡層間的預測準確度與臨床決策之公平性。 方法：本研究提出一套異質集成學習架構，整合多源臨床數據。系統應用自然語言處理技術擷取非結構化臨床文本中之語意特徵，並結合深度卷積神經網路萃取醫學影像之型態特徵。模型建構階段採用基因演算法（GA）最佳化 XGBoost、SVM 與 Elastic Net 等基分類器之權重分配。此外，本研究導入年齡專一性動態閾值機制作為後處理步驟，根據不同年齡層重新校準決策邊界。 結果：在異質臨床資料集之評估中，本研究所提出之多模態集成系統展現了穩定的診斷效能，其曲線下面積（AUC）達 0.8967，整體敏感度達 87.1%。相較於靜態閾值系統，導入動態閾值機制不僅降低了年輕族群的偽陰性率，亦有效控制高齡族群之偽陽性率。 結論：本研究建構之 CDSS 有效改善了單一模態架構與固定決策邊界之侷限。透過提升模型預測力並確保各年齡層間之評估公平性，本系統具備輔助婦產腫瘤領域臨床醫師進行術前診斷之應用潛力。

Background: Ovarian cancer is characterized by high mortality due to insidious onset and a lack of specific early-stage symptoms. Conventional diagnosis relies heavily on single modalities like CA-125 or ultrasonography, which often suffer from suboptimal specificity and subjective bias. Furthermore, conventional systems rely on a static, universal diagnostic threshold, introducing significant age-specific confounding bias between pre- and post-menopausal cohorts. Objective: This research aims to design, implement, and evaluate a comprehensive, high-performance Multi-modal Clinical Decision Support System (CDSS) for preoperative ovarian cancer diagnosis that enhances both predictive accuracy and clinical equity across different age groups. Methods: We proposed a heterogeneous ensemble learning framework that synergistically integrates multi-source clinical data. Specifically, advanced natural language processing (Multilingual-E5-Large) was deployed to capture semantic nuances from unstructured clinical texts, while a deep convolutional network (DenseNet-121) extracted macro-morphological features from medical imaging. A Genetic Algorithm (GA) was utilized to mathematically optimize the weight allocation of distinct base classifiers, including XGBoost, SVM, and Elastic Net. Crucially, an innovative age-specific adaptive dynamic thresholding mechanism was engineered as a post-processing step to recalibrate decision boundaries based on individual age cohorts. Results: Evaluated on a heterogeneous clinical dataset, the proposed multi-modal ensemble system demonstrated superior diagnostic performance, achieving an Area Under the Curve (AUC) of 0.8967 and a sensitivity of 87.1%. Compared to static threshold systems, incorporating the dynamic thresholding mechanism significantly reduced the false-negative rate in younger cohorts and effectively controlled the false-positive rate in older demographics. Conclusion: The proposed CDSS effectively addresses the limitations of single-modality frameworks and fixed decision boundaries. By improving robust predictive performance and ensuring age-specific clinical equity, this system demonstrates the potential to serve as a reliable and clinically actionable decision support tool for practitioners in gynecologic oncology.