臨床應用機器學習於退化性頸椎疾病之影像判讀

Abstract

背景 退化性頸椎疾病是常見的老化現象，主要影響中壯年人，可能導致頸部、肩膀或手臂的疼痛、麻木與肌力下降，嚴重者甚至會失能，對個人生活與社會生產力造成影響。影像學檢查是診斷的關鍵工具，X-ray能顯示骨性變化但難以準確評估神經壓迫，MRI雖能清楚呈現椎間盤與神經結構，但受限於醫療資源與長時間等待，可能延誤治療時機，導致病情惡化。CT在骨性病變診斷上具一定準確性，但對軟組織敏感度不及MRI。依據目前研究顯示人工智慧（AI）影像分析逐漸成為改善診斷效率與準確性的潛在解決方案，有望減少人為誤判與診斷延誤，並提升醫療資源的運用效率。 實驗方法 本研究為回溯性病歷研究，分析2012年1月至2016年12月間在新光醫院因頸椎疾病就診之803位年齡介於18至80歲之患者，收集其X-ray、CT與MRI影像資料，並應用深度卷積神經網路中的YOLOv3模型進行頸椎病變影像的自動化偵測與預測。本研究已獲新光醫院倫理審查委員會核准（IRB: 20220121R）。模型訓練聚焦於分析患者的頸椎側面X-ray影像，針對C3至C7椎體進行病變判讀，包括椎體變形、骨刺形成及鈣化等特徵，藉由影像資料標註與模型訓練建立能辨識有臨床意義的退化性病灶之AI系統，作為臨床決策之輔助依據。 結果 本研究所訓練之機器學習模型於頸椎影像病灶偵測上展現中上程度的分類效能，整體AUC值為0.78，95%信賴區間介於0.73至0.83，且P值小於0.001，與隨機預測相比具有顯著統計差異，顯示模型在病灶辨識上具實用價值。就模型於不同頸椎節段之表現而言，皆展現出一定程度的判別力，能有效從影像中辨認出可能存在病變的位置與類型。該結果支持AI模型在X光影像中進行初步篩檢或輔助診斷的潛力，可提升醫療判讀效率並減少人為疏漏。 結論 本研究驗證了深度學習模型於退化性頸椎病變影像判讀上的可行性，特別是在頸椎後韌帶鈣化物偵測與病灶節段定位方面展現不錯效能，未來具作為輔助診斷工具的潛力。惟本研究仍受限於樣本數量、影像標註品質與模型泛化能力等因素，未來可透過擴增資料量、導入可解釋性AI、多模態融合與強化學習等技術加以改善，並進一步評估其在不同醫療環境下之實用性與成本效益。本研究證實AI模型可有效辨識頸椎影像中之關鍵病變特徵，並提供具參考價值之標註資訊，為AI輔助診斷於脊椎醫學領域奠定實務應用基礎。

Background Degenerative cervical spine disease is a common age-related condition predominantly affecting middle-aged adults, potentially leading to neck, shoulder, or arm pain, numbness, muscle weakness, and even disability. These impairments may compromise individuals’ quality of life and societal productivity. Imaging plays a crucial role in diagnosis: X-rays can reveal bony changes but have limited capacity to assess neural compression; MRI offers detailed visualization of soft tissue and neural structures but is often hindered by long wait times due to limited resources; CT provides good resolution for bone but lacks sensitivity for soft tissue. Recent studies suggest that artificial intelligence (AI)-based imaging analysis may enhance diagnostic efficiency and accuracy, reducing misinterpretation and delays, while optimizing resource utilization. Methods This retrospective study analyzed imaging data from 803 patients aged 18 to 80 years who presented with cervical spine disorders at Shin Kong Hospital between January 2012 and December 2016. Cervical spine X-ray, CT, and MRI images were collected. A deep convolutional neural network using the YOLOv3 architecture was trained to automatically detect and predict degenerative changes on lateral cervical spine X-rays, specifically targeting the C3 to C7 vertebral levels. The model was designed to identify clinically significant abnormalities such as vertebral deformities, osteophyte formation, and calcification. Institutional review board approval was obtained (IRB: 20220121R).

Results The developed AI model demonstrated moderate-to-good diagnostic performance, achieving an overall area under the curve (AUC) of 0.78 (95% CI: 0.73–0.83; P < 0.001), indicating statistically significant improvement over random prediction. The model effectively identified potential pathological segments across C3–C7 and showed promising classification capabilities. These findings support the potential utility of AI models as preliminary screening or assistive diagnostic tools for cervical spine disorders. Conclusion This study confirmed the feasibility of using deep learning models for interpreting degenerative cervical spine changes on X-ray images. The model showed favorable performance in detecting posterior longitudinal ligament calcification and in localizing affected vertebral segments, with potential to assist clinical decision-making. Limitations include the sample size, annotation quality, and generalizability of the model. Future improvements may involve expanding data sets, incorporating explainable AI techniques, multimodal integration, and reinforcement learning. Overall, the findings highlight the clinical applicability of AI-assisted diagnosis in spinal medicine and establish a foundation for further development in this domain.