基於深度學習之雙視角脊椎X光片壓迫性骨折診斷

Abstract

脊椎壓迫性骨折（Vertebral Compression Fractures, VCFs）是骨質疏鬆常見的併發症，且在臨床診斷中經常被低估，可能導致延誤治療並加重病患預後。雖然以椎體為單位的分類模型在自動化 VCF 偵測上展現潛力，但若直接應用於以病人為單位的診斷，常會導致較高的假陽性率。為了解決此問題，本研究提出一個兩階段的深度學習框架，專門用於偵測腰椎 X 光影像中的 VCF。第一階段進行影像層級的分類，用以過濾正常個案，有效降低假陽性數量，並減少不必要的椎體分析。第二階段針對分割後的單節椎體，結合影像特徵與根據放射科醫師經驗而設計的特徵進行椎體層級的 VCF 診斷，同時使用 Grad-CAM 達到模型預測的可解釋性。我們使用來自台大醫院（NTUH）的大型腰椎 X 光資料集進行評估，該資料集包含數千張的 X 光脊椎影像。我們提出的兩階段系統在病人層級達到 87.0% 準確率、 87.3% 特異度以及85.7% 靈敏度；在椎體層級則達到 97.1% 準確率、97.9% 特異度與 77.3% 靈敏度。實驗結果顯示，兩階段設計不僅提升診斷準確性，亦能提供臨床可解釋的預測結果，有助於放射科醫師進行早期 VCF 偵測。

Vertebral compression fractures (VCFs) are a common consequence of osteoporosis and are often underdiagnosed in clinical settings, potentially leading to delayed treatment and worsened patient outcomes. While vertebra-level classification models have shown effectiveness in automated VCF detection, they often suffer from high false positive rates when applied directly to patient-level diagnosis. To address this issue, we propose a two-stage deep learning–based framework for detecting VCFs in lumbar spine X-ray images. The first stage performs image-level classification to filter out normal cases, effectively reducing the number of false positives and minimizing unnecessary vertebra-level evaluations. The second stage focuses on vertebra-level diagnosis by analyzing segmented vertebral bodies using both image and features based on radiologist insights. Grad-CAM is employed to enhance the interpretability of model predictions. We evaluated our method using a dataset from National Taiwan University Hospital (NTUH), which contains thousands of lumbar spine X-rays. The proposed two-stage system achieved 87.0% accuracy, 87.3% specificity, and 85.7% sensitivity at the patient level and 97.1% accuracy, 97.9% specificity, and 77.3% sensitivity at the vertebra level. These results demonstrate that our two-stage design not only improves diagnostic precision but also offers clinically interpretable predictions, making it a valuable tool for aiding radiologists in early VCF detection.