應用半監督深度學習與影像增強於髖關節發育不良超音波影像之分割研究

研究目的：隨著深度學習技術的進步，電腦輔助診斷（Computer-Aided Diagnosis, CAD）已成為醫學影像辨識與分析領域的重要趨勢，特別是在提升臨床診斷效率與減輕醫師工作負擔方面，展現出極大的應用潛力。醫學影像分割技術在辨識器官、病灶及解剖結構中扮演關鍵角色，但超音波影像由於成像不穩定及影像品質變異大，導致分割任務面臨諸多技術挑戰。此外，醫學影像標註工作耗時且成本高昂，致使可用的標註數據相對有限，進一步影響分割模型的表現與臨床應用的普及。發育性髖關節發育不良（Developmental Dysplasia of the Hip, DDH），又稱為先天性髖關節脫臼（Congenital Dislocation of the Hip, CDH），是一種新生兒及嬰幼兒常見的髖關節發育異常疾病。部分患者在新生兒階段即能被診斷，但亦有部分病例在成長過程中逐步顯現症狀。DDH不必然表現為完全脫臼，部分病例僅為髖臼對股骨頭覆蓋不足。臨床上，DDH的早期篩檢通常於嬰兒出生後兩至三個月進行，常用的臨床體檢方法包括Ortolani Test（復位測試）、Barlow Test（脫位測試）、Allis Test及Telescope Sign等。由於新生兒股骨頭尚未鈣化，傳統X光檢查靈敏度有限，因此超音波檢查成為主要的診斷工具，協助評估髖臼發育與股骨頭相對位置。隨著嬰兒年齡增長，股骨頭活動度降低，臨床體檢的適用性亦受影響，此時診斷需結合超音波及X光影像，並觀察雙側大腿外展的對稱性。若DDH未及時發現並矯治，於學步期後症狀將更明顯，單側脫臼可能導致跛行，雙側則可能出現鴨子步態（waddling gait）等步態異常。研究方法及資料：本研究針對超音波影像標記數據有限的困境，採用多重資料增強（Data Augmentation）技術以提升模型的訓練效能與泛化能力。有標記資料部分，採用旋轉、縮放、平移、影像增強等影像增強方法，生成多樣化的訓練樣本，同時維持標記分布的合理性；而對於無標記資料，則透過多個影像增強策略，創建能保留原始特徵且具備額外資訊的數據樣本，以擴展模型的學習空間。本研究採用半監督式深度學習（Semi-Supervised Deep Learning）方法進行影像分割，透過少量標記數據學習影像中的關鍵特徵，從而降低對標記數據的需求並提升分割準確度。在模型設計上，參考近年來不確定性估計（Uncertainty Estimation）及不確定性加權最小化約束（Uncertainty-Weighted Loss Minimization）等方法，利用Kullback-Leibler（KL）散度衡量多尺度預測結果間的不一致性，以單次前向傳遞（forward pass）計算不確定性，兼顧計算效率與模型性能。 本研究所用數據集包含6075張來自多家臨床醫療機構的DDH超音波影像，兼具病例多樣性。針對影像前處理，實驗採用全域直方圖均衡化（GHE）、對比度限制自適應直方圖均衡化（CLAHE）等方法，提升影像可視性及解剖結構辨識度。研究結果： 實驗結果顯示，本研究所提出的方法在多項評估指標上均優於現有對照模型。尤其在Dice指數、準確率及交集並集比（IoU）等指標下，模型展現出更高的影像分割準確性與邊界重建能力。具體而言，在標記數據有限的情境下，本研究所設計的半監督式深度學習模型能有效提升分割任務的表現，並對臨床應用具高度潛力。結論與建議： 本研究證實，結合資料增強與半監督式深度學習方法，有助於解決醫學影像標記數據不足下的分割挑戰，並顯著提升超音波影像於髖關節發育不良診斷的自動化程度。未來建議可持續探索更先進的資料增強技術與生成模型，以進一步豐富訓練數據的多樣性，並結合強化學習或遷移學習等先進機器學習方法，發展更具專屬性的模型架構，以因應不同類型醫學影像的需求。隨著運算資源與臨床數據的進一步累積，實時醫學影像分析與個人化診斷系統的應用可望更趨成熟，最終能提升臨床診斷效率與準確性，並且推動精準醫療與智慧健康照護。

Objective：With the rapid advancement of deep learning, computer-aided diagnosis (CAD) has emerged as a focal point in medical image recognition and analysis. Among its many applications, medical image segmentation plays a crucial role in identifying organs, lesions, and anatomical structures. The developing of accurate segmentation techniques not only enhances clinical decision-making efficiency but also substantially alleviates the workload of healthcare professionals. However, ultrasound image segmentation presents significant challenges due to inherent imaging instability and variations in image quality. Additionally, the labor-intensive annotation process typically results in a limited amount of labeled data, which restricts model performance and clinical applicability.Developmental Dysplasia of the Hip (DDH), formerly known as Congenital Dislocation of the Hip (CDH), is a condition characterized by abnormal development of the hip joint. While some cases can be identified in the neonatal period, become apparent later during infancy. Early diagnosis, often performed within two to three months after birth, relies on physical examinations such as the Ortolani Test, Barlow Test, Allis Test, and Telescope Sign. Because the femoral head is not yet ossified in early infancy, ultrasound imaging is preferred over X-rays for assessing acetabular development and femoral head positioning. After three months, as hip mobility decreases, diagnosis increasingly depends on asymmetry in hip abduction, in combination with ultrasound and radiographic assessment. In toddlers, DDH symptoms become more pronounced; unilateral dislocation may cause limping, while bilateral dislocation often results in a waddling gait. Methods and Materials：To address the challenges of limited labeled data, this study employs comprehensive data augmentation techniques to improve model training efficacy and generalization capability. For labeled data, augmentation methods such as rotation, scaling, and translation are utilized to create diverse training samples that maintain annotation integrity. For unlabeled data, generative models and fusion strategies are leveraged to produce new samples that retain essential characteristics of the original images while introducing additional informative features.A semi-supervised deep learning approach framework is adopted, enabling the model to learn critical features from a small set of labeled images and a larger set of unlabeled images, thus reducing dependence on extensive manual annotation and enhancing segmentation accuracy. This approach incorporates uncertainty estimation and uncertainty-weighted loss minimization, following recent research. Uncertainty estimation is implemented by measuring inconsistencies across multi-scale predictions using Kullback-Leibler (KL) divergence, providing a computationally efficient alternative to traditional methods such as Monte Carlo Dropout. The experimental design includes the dataset composition, ultrasound image enhancement strategies, data augmentation techniques, model architecture, and evaluation metrics. The dataset consists of 6,075 ultrasound images collected from multiple clinical sources, ensuring a diverse patient population. Image enhancement methods such as Global Histogram Equalization (GHE) and Contrast-Limited Adaptive Histogram Equalization (CLAHE) are employed to improve image visibility and anatomical detail recognition.Results：Experimental results demonstrate that the proposed method surpasses comparative models across multiple evaluation metrics. Notably, the model achieves superior image segmentation accuracy and boundary reconstruction, as evidenced by higher Dice coefficient, accuracy, and Intersection over Union (IoU) scores. These results underscore the efficacy and proposed approach in medical image segmentation tasks.Conclusion and Suggestion：This study confirms that combining data augmentation with semi-supervised deep learning can effectively address the limitations imposed by scarce labeled data in medical image segmentation, particularly for DDH ultrasound images. Future research should explore advanced augmentation and generative modeling techniques to further increase training data diversity and model robustness. Integrating deep learning with other machine learning paradigms, such as reinforcement learning or transfer learning, may yield further improvements in medical image analysis. Additionally, the development of specialized architectures tailored to different types of medical images and diagnostic requirements is recommended. As computational resources continue to evolve, real-time applications for medical image analysis are expected to become increasingly practical, thereby improving clinical diagnostic efficiency and accuracy. These advancements will foster the development of personalized medicine, providing more precise and effective diagnostic and therapeutic strategies for patients.