van der Gaast N, Bagave P, Assink N, Broos S, Jaarsma R L, Edwards M J R, Hermans E, IJpma F F A, Ding A Y, Doornberg J N, Oosterhoff J H F
Department of Orthopaedics & Trauma Surgery, Flinders Medical Centre and Flinders University, Adelaide, SA, Australia; Department of Trauma Surgery, Radboud University Medical Center, Radboud University Nijmegen, the Netherlands.
Department of Engineering Systems and Services, Faculty of Technology Policy and Management, Delft University of Technology, Delft, the Netherlands.
Knee. 2025 Jun;54:81-89. doi: 10.1016/j.knee.2025.02.001. Epub 2025 Mar 1.
Deep learning (DL) has been shown to be successful in interpreting radiographs and aiding in fracture detection and classification. However, no study has aimed to develop a computer vision model for tibia plateau fractures using the Schatzker classification. Therefore, this study aims to develop a deep learning model for (1) detection of tibial plateau fractures and (2) classification according to the Schatzker classification.
A multicenter approach was performed for the collection of radiographs of patients with tibia plateau fractures. Both anteroposterior and lateral images were uploaded into an annotation software and manually labelled and annotated. The dataset was balanced for optimizing model development and split into a training set and a test set. We trained two convolutional neural networks (GoogleNet and ResNet) for the detection and classification of tibia plateau fractures following the Schatzker classification.
A total of 1506 knee radiographs from 753 patients, including 368 tibial plateau fractures and 385 healthy knees, were used to create the algorithm. The GoogleNet algorithm demonstrated high sensitivity (92.7%) but intermediate accuracy (70.4%) and positive predictive value (64.4%) in detecting tibial plateau fractures, indicating reliable detection of fractured cases. It exhibited limited success in accurately classifying fractures according to the Schatzker system, achieving an accuracy of only 34.6% and a sensitivity of 32.1%.
This study shows that detection of tibial plateau fractures is a task that a DL algorithm can grasp; further refinement is necessary to enhance their accuracy in fracture classification. Computer vision models might improve using different classification systems, as the current Schatzker classification suffers from a low interobserver agreement on conventional radiographs.
深度学习（DL）已被证明在解读X线片以及辅助骨折检测和分类方面取得了成功。然而，尚无研究旨在开发一种基于Schatzker分类的胫骨平台骨折计算机视觉模型。因此，本研究旨在开发一种深度学习模型，用于（1）检测胫骨平台骨折和（2）根据Schatzker分类进行分类。
采用多中心方法收集胫骨平台骨折患者的X线片。前后位和侧位图像均上传至注释软件并进行手动标记和注释。对数据集进行平衡以优化模型开发，并分为训练集和测试集。我们训练了两个卷积神经网络（GoogleNet和ResNet），用于按照Schatzker分类对胫骨平台骨折进行检测和分类。
共使用了来自753例患者的1506张膝关节X线片，其中包括368例胫骨平台骨折和385例健康膝关节，以创建该算法。GoogleNet算法在检测胫骨平台骨折时显示出高灵敏度（92.7%），但准确度中等（70.4%），阳性预测值为（64.4%），表明对骨折病例的检测可靠。在根据Schatzker系统准确分类骨折方面，它的成功率有限，准确度仅为34.6%，灵敏度为32.1%。
本研究表明，检测胫骨平台骨折是DL算法能够掌握的一项任务；需要进一步优化以提高其在骨折分类方面的准确性。计算机视觉模型可能会因使用不同的分类系统而得到改进，因为目前的Schatzker分类在传统X线片上观察者间一致性较低。
