針對新興流行病的雙路徑肺部超音波分析：基於FocalNet-AE的無監督異常偵測與少樣本分類

Abstract

在新興呼吸道傳染病爆發的關鍵初期，臨床應對常受限於確診病例稀缺與影像標註成本高昂，致使傳統深度學習模型陷入數據不足的冷啟動(Cold-start)困境。肺部超音波(Lung ultrasound)憑藉其高可攜性、無游離輻射及即時床邊檢測之優勢，成為理想的第一線篩檢工具，惟其高度依賴操作者經驗及顯著的影像變異性，為自動化分析帶來巨大挑戰。為解決上述限制，本研究提出一套創新的雙路徑分析架構，旨在有限數據條件下提供穩定的分析工具。核心模型採用FocalNet自動編碼器(Autoencoder)作為骨幹網路，並引入遮罩自動編碼器(Mask Autoencoder)訓練策略與混合損失函數，於超過30,000張無標註醫療超音波影像上進行預訓練，以學習具備高度結構判別力的穩健特徵表示。在此架構下，本研究整合了兩大技術流程，首先是無監督異常偵測流程，該流程利用滑動視窗計算區域重建誤差，並結合統計閾值機制，在完全無須病理標註的情況下即可識別異常影像，於COVID-19數據集上達到了0.770的平衡準確率。其次，針對臨床隨後逐漸取得少量確診病例的情境，本研究設計了少樣本學習(few-shot learning)流程，利用預訓練編碼器作為強效特徵提取器，僅需微調單層線性分類器即可快速適應特定疾病。實驗數據顯示，僅需3%的極少量標註數據，模型即可達成高達0.923的平衡準確率，展現優異的數據效益與臨床實用性。總結而言，本研究建立了一套具可擴展性的工程解決方案，實現了從無監督偵測到少樣本監督學習的漸進式過渡，有效克服了新興傳染病初期的數據稀缺難題，為未來面對大流行病時的快速反應與輔助篩檢，提供了一項具備高度臨床價值的技術架構。

During the critical early stages of emerging respiratory infectious disease outbreaks, the clinical response is often hampered by a scarcity of confirmed cases and the high costs associated with annotation. This makes conventional supervised deep learning models less effective due to the "cold-start" dilemma. Lung ultrasound (LUS) is an ideal first-line screening tool given its portability, absence of ionizing radiation, and real-time bedside capability. However, its effective deployment is hindered by significant operator dependency and considerable image variability. To address these challenges, this study introduces a novel Dual-Pipeline Framework based on a FocalNet Autoencoder (FocalNet-AE) that integrates unsupervised anomaly detection with few-shot supervised classification capabilities. The core model was pre-trained on over 30,000 unlabeled medical ultrasound images using masked autoencoding combined with a hybrid loss function to learn robust structural representations. Within this framework, the unsupervised pipeline employs sliding-window reconstruction error analysis to identify anomalies without prior knowledge of pathology, achieving a balanced accuracy of 0.770 in zero-shot scenarios. Additionally, the supervised pipeline utilizes the pre-trained encoder to swiftly adapt to specific diseases with minimal labels, achieving a balanced accuracy of 0.923 using only 3% of the training data. This study presents a data-efficient and scalable engineering solution that facilitates a smooth transition from unsupervised anomaly detection to few-shot classification, providing a clinically applicable tool for rapid response during future pandemics.