组织工程与重建外科杂志

组织工程与重建外科杂志 ›› 2024, Vol. 20 ›› Issue (1): 114-.

• • 上一篇    下一篇

深度学习系统辅助成人发育性髋关节发育不良的分型培训

  

  • 发布日期:2024-03-07

Deep-learning system assisted staged training for developmental dysplasia of the hip

  • Published:2024-03-07

PDF (PC)

115

摘要：

　目的　开发一种深度学习系统用于成人发育性髋关节发育不良（Developmental dysplasia of the hip，DDH）患者
的Crowe分型辅助诊断，并且分析该系统对于帮助临床医学生掌握DDH分型的可行性。方法　纳入149例X线片训练
集、42例测试集以及21例验证集，分割盆骨、提取DDH局部图像块，将金标准结果与医学生、AI辅助医学生评估结果进
行比较。结果　测试集共纳入42例，其中女性30例，男性12例，年龄（69±12）岁，涉及发育不良髋关节67侧（左30侧，
右 37 侧）。AI、医学生、AI 辅助医学生评估结果与金标准的相关性为 0.906［95% CI（0.850，0.941）］、0.823［95% CI
（0.726，0.887）］、0.886［95% CI（0.821，0.929）］；准确率分别为0.87、0.78、0.88；精确度分别为0.88、0.83、0.89；召回率分别为 0.87、0.78、0.88；F1 值分别为 0.87、0.80、0.88。混淆矩阵和条件概率结果显示，预测准确率Ⅰ型 DDH 三组分别为
0.98、0.88、0.96，Ⅱ型 DDH 三组分别为 0.40、0.20、0.40，Ⅲ型 DDH 三组分别为 0.56、0.67、0.78；Ⅳ型 DDH 三组分别为0.88、0.75、0.88。结论　深度学习辅助诊断系统可以有效提高医学生对于DDH分型的评估能力，可作为医学生学习掌
握DDH影像诊断的培训工具。

关键词:

Abstract:

Objective To develop a deep learning system for assisted diagnosis of Crowe staging in adult patients with
developmental dysplasia of the hip （DDH）， and to analyze the feasibility of the system in assisting clinical medical students
to master DDH staging. Methods A training set of 149 X-rays， a test set of 42 cases， and a validation set of 21 cases were
included， and the pelvis was segmented， localized image blocks of DDH were extracted， and the gold-standard results were
compared with those assessed by medical students and AI-assisted medical students. Results A total of 42 cases， including
30 females and 12 males， aged （69±12） years， were included in the test set， and 67 dysplastic hips were involved （30 on the left and 37 on the right） . The correlation of the AI， medical student， and AI-assisted medical student assessment results with the gold standard was 0.906［ 95% CI（ 0.850， 0.941）］， 0.823［ 95% CI（ 0.726， 0.887）］， 0.886［ 95% CI（0.821， 0.929）］ . The accuracy of AI， medical students and AI-assisted medical students was 0.87，
0.78 and 0.88， the precision was 0.88，0.83 and 0.89， the recall rate was 0.87，0.78 and 0.88， and F1 value was 0.87，0.80 and 0.88， respectively. The results ofthe confusion matrices and conditional probabilities showed that the accuracy of the three groups of type Ⅰ were 0.98，0.88，0.96， and 0.40，0.20，0.40 for type Ⅱ trio， and 0.56，0.67，0.78 for type Ⅲ trio， and 0.88，0.75，0.88 for type Ⅳ trio.
Conclusion Deep learning-assisted diagnostic system can effectively improve the medical students' assessment of the DDH
patients with various types of DDH， and can be used as a training tool for medical students to learn and master the diagnosis
of DDH imaging.

Key words:

medical students