| 摘 要: 在膝关节骨折检测任务中,传统的诊断方法因诊断效率和准确性存在局限性,通过引入膝关节骨折的自动化诊断与分析,可以显著提升检测精度。为应对低对比度和复杂结构的挑战,提出一种改进的YOLOv8算法,结合ESFAM多尺度特征聚合模块、ARAH自适应区域注意力检测头和BDC-Conv双通道膨胀卷积模块,优化了模型的检测性能。通过实验验证,改进后的模型在6种膝关节骨折类型的识别上表现出更高的检测准确性,且在召回率和平均精度(mAP)方面均优于标准YOLOv8算法。该研究为膝关节骨折的智能检测提供了有效的技术方案和方法支持。 |
| 关键词: 膝关节骨折 YOLOv8 多尺度特征聚合 注意力检测头 |
|
中图分类号: TP183
文献标识码: A
|
| 基金项目: 贵州省基础研究计划项目(自然科学)(黔科合基础-ZK[2023]一般052) |
|
| Intelligent Analysis Algorithm and Model Construction for Knee Joint Medical Images Based on Neural Networks |
|
DU Changlun, WU Peng, REN Jialong, ZHANG Jing, LIU Yongxing
|
(State Key Laboratory of Public Big Data Jointly Built by the Provincial and Ministerial Government, Guizhou University, Guiyang 550025, China)
528752329@qq.com; pengwu@gzu.edu.cn; pigo9216@163.com; 906351319@qq.com; 544073227@qq.com
|
| Abstract: In orthopedic imaging, automated diagnosis and analysis of knee joint fractures are crucial for addressing the limitations of traditional diagnostic methods, with the goal of significantly enhancing detection accuracy.To tackle the challenges of low contrast and complex structures, this study proposes an improved YOLOv8 algorithm. It incorporates an ESFAM (Efficient Scale-wise Feature Aggregation Module) for multi- scale feature aggregation, an ARAH (Adaptive Region-Aware Attention Head) for adaptive region-focused detection, and a BDC-Conv (Bi-Directional Dilated Convolution) module. These enhancements optimize the model’s detection performance. Experimental validation demonstrates that the improved model achieves higher detection accuracy in identifying six types of knee joint fractures. It also outperforms the standard YOLOv8 algorithm in both recall rate and mean Average Precision (mAP).This research provides a more effective technical solution and methodological support for the intelligent detection of knee joint fractures. |
| Keywords: knee joint fractures YOLOv8 mult-i scale feature aggregation attention detection head |
