Aims The purpose of this study was to develop a convolutional neural network (CNN) for fracture detection, classification, and identification of greater tuberosity displacement ≥ 1 cm, neck-shaft angle (NSA) ≤ 100°, shaft translation, and articular fracture involvement, on plain radiographs. Methods The CNN was trained and tested on radiographs sourced from 11 hospitals in Australia and externally validated on radiographs from the Netherlands. Each radiograph was paired with corresponding CT scans to serve as the reference standard based on dual independent evaluation by trained researchers and attending orthopaedic surgeons. Presence of a fracture, classification (non- to minimally displaced; two-part, multipart, and glenohumeral dislocation), and four characteristics were determined on 2D and 3D CT scans and subsequently allocated to each series of radiographs. Fracture characteristics included greater tuberosity displacement ≥ 1 cm, NSA ≤ 100°, shaft translation (0% to < 75%, 75% to 95%, > 95%), and the extent of articular involvement (0% to < 15%, 15% to 35%, or > 35%). Results For detection and classification, the algorithm was trained on 1,709 radiographs (n = 803), tested on 567 radiographs (n = 244), and subsequently externally validated on 535 radiographs (n = 227). For characterization, healthy shoulders and glenohumeral dislocation were excluded. The overall accuracy for fracture detection was 94% (area under the receiver operating characteristic curve (AUC) = 0.98) and for classification 78% (AUC 0.68 to 0.93). Accuracy to detect greater tuberosity fracture displacement ≥ 1 cm was 35.0% (AUC 0.57). The CNN did not recognize NSAs ≤ 100° (AUC 0.42), nor fractures with ≥ 75% shaft translation (AUC 0.51 to 0.53), or with ≥ 15% articular involvement (AUC 0.48 to 0.49). For all objectives, the model's performance on the external dataset showed similar accuracy levels. Conclusion CNNs proficiently rule out proximal humerus fractures on plain radiographs. Despite rigorous training methodology based on CT imaging with multi-rater consensus to serve as the reference standard, artificial intelligence-driven classification is insufficient for clinical implementation. The CNN exhibited poor diagnostic ability to detect greater tuberosity displacement ≥ 1 cm and failed to identify NSAs ≤ 100°, shaft translations, or articular fractures.

中文翻译：

---

X 线平片上肱骨近端骨折的检测、分类和特征。


目的 本研究的目的是开发一个卷积神经网络 （CNN），用于骨折检测、分类和识别 1 cm ≥大粗隆移位、颈干角 （NSA） ≤ 100°、轴平移和关节骨折受累，在 X 线平片上。方法 CNN 在来自澳大利亚 11 家医院的 X 光片上进行了训练和测试，并在来自荷兰的 X 光片上进行了外部验证。每张 X 光片与相应的 CT 扫描配对，作为基于训练有素的研究人员和主治骨科医生双重独立评估的参考标准。在 2D 和 3D CT 扫描中确定骨折、分类 （非至轻微移位;两部分、多部分和盂肱关节脱位）和 4 个特征，随后分配给每个系列的 X 光片。骨折特征包括较大的粗隆移位≥ 1 cm、NSA ≤ 100°、骨干平移（0% 至 < 75%、75% 至 95%、> 95%）和关节受累程度（0% 至 < 15%、15% 至 35% 或 > 35%）。结果对于检测和分类，该算法在 1,709 张 X 光片 （n = 803） 上进行了训练，在 567 张 X 光片上进行了测试 （n = 244），随后在 535 张 X 光片上进行了外部验证 （n = 227）。对于特征描述，健康肩部和盂肱关节脱位被排除在外。骨折检测的总体准确性为 94% （受试者工作特征曲线下面积 （AUC） = 0.98），分类的总体准确性为 78% （AUC 0.68 至 0.93）。检测 1 cm ≥较大粗隆突骨折移位的准确性为 35.0% （AUC 0.57）。CNN 未识别 100° ≤ （AUC 0.42） 的 NSA，也未识别轴平移≥ 75% 的骨折（AUC 0.51 至 0.53），或关节受累≥ 15%（AUC 0.48-0.49）。对于所有目标，模型在外部数据集上的性能都显示出相似的准确性水平。结论 CNN 在 X 线平片上可熟练排除肱骨近端骨折。尽管以 CT 成像为基础的严格训练方法以多评分者共识作为参考标准，但人工智能驱动的分类不足以进行临床实施。CNN 检测 1 cm ≥较大的结节移位的诊断能力较差，并且无法识别 100° ≤ NSA、干平移或关节骨折。