基於深度學習的下肢X射線骨骼關鍵點和姿態測量

Abstract

背景： 醫學影像的量化分析在骨骼肌肉放射科醫學領域至關重要，但傳統手動測量方法耗時、費力且易受主觀性影響，導致結果變異性大。深度學習在一般影像關鍵點偵測上表現卓越，然而直接應用於醫學X射線影像仍面臨獨特挑戰，例如金屬植入物造成的遮擋和影像品質不佳等問題。 目的： 本研究旨在開發一個高效、客觀且高精度的自動化工具，透過深度學習模型實現下肢骨骼關鍵點的自動偵測與姿態的精確量測，以解決臨床上人工測量的痛點，提升診斷的標準化與精準度。 方法： 研究採用深度學習中基於直接關鍵點標註的策略，從全下肢站立位X射線影像中預測髖關節、膝關節和踝關節的精確像素坐標。資料集包含來自2897位病患的4396張長腿X射線影像，並採用平衡劃分策略，考量了金屬植入物等特徵的分佈，以確保模型的代表性與穩健性。模型的效能透過訓練損失、交叉驗證、預測與真實值比較視覺化、關鍵點坐標預測誤差分析及Bland-Altman分析進行評估。 結果： 本研究對深度學習模型進行了全面評估。訓練損失與交叉驗證結果顯示模型穩健且準確。模型在關鍵點座標預測誤差方面表現優異，廣義線性模型分析揭示患者年齡及性別與植入物有無的交互作用對誤差有顯著影響。Bland-Altman分析顯示髖膝踝角（HKA）的模型預測值與真實值之間偏差為 0.0197 度；髖踝距（HA）的偏差為 0.136 毫米。 結論： 本研究成功開發基於YOLOv11 Pose的下肢X射線骨骼關鍵點自動化偵測與姿態測量系統，其中YOLO11l-pose模型表現最優。為提升泛用性，本研究納入多樣化影像並分析座標誤差的特徵貢獻度，發現在平衡資料集劃分下，能達到單獨的性別、身高、體重、BMI及植入物有無則無顯著影響。儘管面臨真實值標註與資料來源限制，此工具仍具備顯著提升骨科診斷標準化與精準度的潛力。

Background: Quantitative analysis of medical images is crucial in musculoskeletal medicine or radiology, but traditional manual measurement methods are time-consuming, labor-intensive, subjective, and prone to variability. While deep learning excels in general image keypoint detection, its direct application to medical X-ray images faces unique challenges, such as obliterations caused by metal implants and suboptimal image quality. Purpose: This study aims to develop an efficient, objective, and highly accurate automated tool using deep learning models for automatic detection of lower limb skeletal keypoints and precise posture measurement from X-ray images, thereby addressing the pain points of manual measurements in clinical practice and enhancing diagnostic standardization and accuracy. Methods: This study employs a deep learning strategy based on direct keypoint annotation to predict the precise pixel coordinates of hip, knee, and ankle joints from full-leg standing X-ray radiographs. The dataset comprises 4396 long-leg X-ray images from 2897 distinct patients, utilizing a balanced splitting strategy that considers the distribution of features like metal implants to ensure model representativeness and robustness. Model performance was evaluated using training loss, cross-validation, visualization of predicted versus ground truth comparisons, keypoint coordinate prediction error analysis, and Bland-Altman analysis. Results: This study comprehensively evaluated the deep learning model. Training loss and cross-validation results indicated model robustness and accuracy. Model demonstrated superior performance in keypoint coordinate prediction error, with Generalized Linear Model analysis revealing that patient age and the interaction between sex and implant presence significantly influenced error. Bland-Altman analysis showed a bias of 0.0197 degrees for Hip-Knee-Ankle (HKA) angle; and a bias of 0.136 millimeters for Hip-Ankle (HA) length. Conclusion: This study successfully developed an automated system for lower limb X-ray skeletal keypoint detection and posture measurement using the YOLOv11 Pose model, with YOLO11l-pose demonstrating superior performance. To enhance generalizability, diverse image data was incorporated, and feature contribution to coordinate errors was analyzed, finding that under balanced dataset division, individual sex, height, weight, BMI, and presence or absence of implants had no significant impact. Despite challenges like ground truth annotation refinement and data source limitations, this tool holds significant potential to enhance orthopedic diagnostic standardization and accuracy.