| 摘 要: 针对医疗设备存储资源有限的问题,提出一种基于复合知识蒸馏的诊断分类方法,旨在确保骨科影像诊断模型的高精度性能。该方法首先采用自适应直方图均衡化对数据集进行增强;其次引入知识蒸馏,选用EfficientNet-B7作为教师网络、EfficientNet-B0作为学生网络,同时对学生网络引入渐进式自蒸馏,以提升特征挖掘和泛化能力。在MURA(Musculoskeletal Radiographs)数据集上进行验证的结果表明,复合知识蒸馏(CKD)模型的准确率为96.26%,其参数量仅为EfficientNet-B7模型参数量的8.48%,并且在准确率方面仅下降了1.16%,验证了此模型的有效性。 |
| 关键词: 骨科影像;自适应直方图均衡化;特征挖掘;知识蒸馏;渐进式自蒸馏 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(51867006);贵州省科学技术计划项目(黔科合支撑[2022]一般264) |
|
| Research on Diagnostic Classification of Orthopedic Images Based on Composite Knowledge Distillation |
|
WANG Kao, WU Qinmu
|
(School of Electrical Engineering, Guizhou University, Guiyang 550025, China)
gzu_xixi@163.com; qmwu@gzu.edu.cn
|
| Abstract: Aiming at the limited storage resources of medical devices, a diagnostic classification method based on Composite Knowledge Distillation(CKD) is proposed to ensure the high-precision performance of orthopedic image diagnostic models. Firstly, adaptive histogram equalization is used to enhance the dataset, and then knowledge distillation is introduced, selecting EfficientNet-B7 as the teacher network and EfficientNet-B0 as the student network. Additionally, an incremental self-distillation approach is applied to the student network to improve feature extraction and generalization capabilities. Validation results on the MURA (Musculoskeletal Radiographs) dataset show that the CKD model achieves an accuracy of 96.26% , with its parameter count being only 8.48% of that of the EfficientNet-B7 model, but there is only a 1.16% decrease in the accuracy, which validates the effectiveness of this model. |
| Keywords: orthopedic image; adaptive histogram equalization; feature mining; knowledge distillation; progressive self-distillation |
