1.安徽建筑大学机电学院(合肥 230061)
2.中国科学院合肥物质科学研究院核能安全技术研究所(合肥 230031)
3.淮北师范大学(安徽淮北 235000)
4.中国科学技术大学附属第一医院(安徽省立医院)骨科(合肥 230001)
张子恒。E-mail:zhzhang@inest.cas.cn
网络首发:2026-08-17,
纸质出版:2026-08-20
移动端阅览
梁自晨,张子恒,张竹岚,等.基于全局分组坐标注意力的数字化放射股骨头坏死分类模型[J].北京生物医学工程,2026,45(4):331-339.
LIANG Zichen,ZHANG Ziheng,ZHANG Zhulan,et al.Digital radiography osteonecrosis classification model based on global grouped coordinate attention[J].Beijing Biomedical Engineering,2026,45(4):331-339.
梁自晨,张子恒,张竹岚,等.基于全局分组坐标注意力的数字化放射股骨头坏死分类模型[J].北京生物医学工程,2026,45(4):331-339. DOI: 10.3969/j.issn.1002-3208.2026.04.001.
LIANG Zichen,ZHANG Ziheng,ZHANG Zhulan,et al.Digital radiography osteonecrosis classification model based on global grouped coordinate attention[J].Beijing Biomedical Engineering,2026,45(4):331-339. DOI: 10.3969/j.issn.1002-3208.2026.04.001.
目的
2
本研究旨在建立一种有效的术后股骨头坏死评估方案,特别关注股骨头数字化放射图像的复杂多样性。股骨头坏死是一种严重的骨科疾病,早期诊断对患者的预后至关重要。传统的影像学评估方法往往依赖于医生的专业判断,容易受到主观因素的影响,且对图像特征的提取和分析能力有限。因此,本文提出了一种基于深度学习的股骨头坏死分类模型,以提高模型对股骨头DR图像识别的准确度及计算效率。
方法
2
本研究采用轻量型网络MobileViT作为主干网络,结合了卷积神经网络(convolutional neural network, CNN)和视觉变换器(Vision Transformer)的优点,以有效处理高维图像数据。为增强模型的特征提取能力,在MV2块前嵌入了Squeeze-and-Excitation Networks(SeNet)注意力机制,通过自适应调整通道权重,帮助模型聚焦于重要特征。此外,课题组设计了全局分组坐标注意力(global grouped coordinate attention, GGCA)模块,使模型能够捕捉多维度的全局信息并进行特征增强。最后,采用基于L1非结构化剪枝策略来降低计算成本,通过去除冗余参数以提高模型的计算效率。
结果
2
本文选取安徽省立医院总院及南区因股骨颈骨折接受闭合复位和空心螺钉复位的470例患者临床资料作为数据集,进行消融实验和对比实验,结果显示该模型的准确率达到94.57%,非零参数量为6.64×10
5
,与改进前相比参数量降低30.14%,与常用CNN模型AlexNet、MobileNet、ShuffleNet相比,准确率分别提高10.24%,16.26%,21.98%。
结论
2
本文提出的基于深度学习的股骨头坏死分类模型在术后股骨头坏死检测中具有较好的性能,是协助医生提高工作效率的一种有效尝试。通过结合轻量型网络架构、注意力机制和剪枝策略,该模型不仅提高了诊断准确性,还降低了计算成本。
Objective
2
This study aims to establish an effective postoperative femoral head necrosis assessment protocol, with a particular focus on the complexity and diversity of digital radiography (DR) images of the femoral head. Femoral head necrosis is a severe orthopedic disease, and early diagnosis is crucial for patient prognosis. Traditional imaging assessment methods often rely on the professional judgment of doctors, are susceptible to subjective factors, and have limited ability to extract and analyze image features. Therefore, this paper proposes a deep learning-based classification model for femoral head necrosis, aiming to improve the accuracy and computational efficiency of the model in recognizing DR images of the femoral head.
Methods
2
This study uses the lightweight network MobileViT as the backbone network, combining the advantages of convolutional neural network (CNN) and vision transformers to effectively process high-dimensional image data. To enhance the feature extraction capability of the model, we embed squeeze and excitation networks (SeNet) attention mechanism before the MV2 block, which helps the model focus on important features by adaptively adjusting channel weights. In addition, we design a global grouped coordinate attention (GGCA) module to enable the model to capture multi-dimensional global information and perform feature enhancement. Finally, an unstructured pruning strategy based on L1 is adopted to reduce computational costs and improve the computational efficiency of the model by removing redundant parameters.
Results
2
This article selects clinical data of 470 patients who received closed reduction and hollow screw reduction for femoral neck fractures in the General Hospital and South District of Anhui Provincial Hospital as the dataset, and conducts ablation experiments and comparative experiments. The results show that the accuracy of the model reaches 94.57%, with a non-zero parameter quantity of 6.64 × 10
5
, which is 30.14% lower than that before the improvement. Compared with the commonly used CNN model AlexNet, MobileNet and ShuffleNet, this model improves accuracy by 10.24%, 16.26%, and 21.98% respectively.
Conclusions
2
The deep learning based classification model for femoral head necrosis proposed in this article has good performance in postoperative detection of femoral head necrosis and is an effective attempt to assist doctors in improving work efficiency. By combining lightweight network architecture, attention mechanism, and pruning strategy, this model not only improves diagnostic accuracy but also reduces computational costs.
石磊, 籍庆余, 陈清威, 等. 视觉Transformer在医学图像分析中的应用研究综述[J]. 计算机工程与应用, 2023, 59(8): 41-55.
Shi L, Ji QY, Chen QW, et al. Review of research on application of vision transformer in medical image analysis[J]. Computer Engineering and Applications, 2023, 59(8): 41-55.
徐锦凡.基于注意力机制与多尺度卷积的医学图像分割方法研究[D]. 桂林:桂林理工大学, 2025.
Xu JF. Research on medical image segmentation methods based on attention mechanism and multi-scale convolution[D]. Guilin: Guilin University of Technology, 2025.
宋宇, 徐睿, 才晓东, 等. 基于多模态融合的3DRes-ViT膝骨关节炎分类模型[J]. 光学精密工程, 2024, 32(21): 3244-3255.
Song Y, Xu R, Cai XD, et al. 3DRes-ViT knee osteoarthritis classification model based on multimodal fusion[J]. Optics and Precision Engineering, 2024, 32(21): 3244-3255.
蒋舒颖, 李志明, 莫贤, 等. 融合注意力的主动迭代优化白细胞图像分类模型[J]. 中国生物医学工程学报, 2024, 43(4): 408-418.
Jiang SY, Li ZM, Mo X, et al. Active iterative optimization of leukocyte image classification model with fused attention[J]. Chinese Journal of Biomedical Engineering, 2024, 43(4): 408-418.
平金如, 孙子文. 特征融合的MV2-Transformer肺炎X光图像分类模型[J]. 计算机应用, 2025, 45(12):4030-4036.
Ping JR, Sun ZW. Pneumonia X-ray image classification model by MV2-Transformer with feature fusion[J]. Journal of Computer Applications, 2025, 45(12): 4030-4036.
周杰.基于深度学习的新冠肺炎CT图像分类[D]. 南京: 南京信息工程大学, 2023.
魏榕剑, 邵剑飞, 温剑, 等. 融合SE和多尺度卷积的轻量级CT图像分类模型[J]. 云南大学学报(自然科学版), 2023, 45(3): 583-589.
Wei RJ, Shao JF, Wen J, et al. A lightweight CT image classification model for fusion SE and multi-scale convolution[J]. Journal of Yunnan University: Natural Sciences Edition, 2023, 45(3): 583-589.
赖裕,李锵,聂为之,等. 基于混合知识蒸馏的轻量级胸部疾病分类算法[J]. 光电子·激光, 2024, 35(9): 993-1000.
Lai Y, Li Q, Nie WZ. Lightweight thoracic disease classification algorithm based on mixed knowledge distillation[J]. Journal of Optoelectronics·Laser, 2024, 35(9): 993-1000.
郑洋,蒋晓天,付东豪,等. 基于剪枝-蒸馏的视觉Transformer模型压缩[J]. 西安电子科技大学学报, 2025, 52(4): 55-65.
Zheng Y, Jiang XT, Fu DH, et al. Vision transformer model compression based on pruning-distillation[J]. 西安电子科技大学学报, 2025, 52(4): 55-65.
苏申申, 周卫, 周淋芋, 等. 基于MobileViT轻量化网络的蘑菇图像分类算法改进[J]. 现代计算机, 2024, 30(21): 69-73.
Su SS, Zhou W, Zhou LY, et al. Improvement of mushroom image classification algorithm based on lightweight MobileViT network[J]. Modern Computer, 2024, 30(21): 69-73.
Tan M, Le Q. Efficientnet: rethinking model scaling for convolutional neural networks[C]// 36th International Conference on Machine Learning. Long Beach, California, USA: PMLR 97, 2019: 6105-6114.
Sinha D, El-Sharkawy M. Thin mobilenet: An enhanced mobilenet architecture[C]//2019 IEEE 10th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON). New York, NY, USA: IEEE Press, 2019: 280-285.
Zhang X, Zhou X, Lin M, et al. Shufflenet: an extremely efficient convolutional neural network for mobile devices[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, NT, USA: IEEE Press, 2018: 6848-6856.
Zhang X. The AlexNet, LeNet-5 and VGG NET applied to CIFAR-10[C]//2021 2nd International Conference on Big Data & Artificial Intelligence & Software Engineering (ICBASE). Nanchang, China: IEEE Press, 2021: 414-419.
李秋宇,李其洋. 基于卷积神经网络的图像分类算法优化[J/OL]. 智能物联技术. [2026-04-14]. https://cnki.net/urlid/33.1411.TP.20260413.1813.008.
Li QY, Li QY. Optimization of image classification algorithm based on convolutional neural network[J/OL]. Technology of IoT & AI. [2026-04-14]. https://cnki.net/urlid/33.1411.TP.20260413.1813.008.
Woo S, Debnath S, Hu R, et al. ConvNeXt V2: Co-designing and scaling ConvNets with masked autoencoders[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver, BC, Canada: IEEE Press, 2023: 16133-16142.
0
浏览量
0
下载量
0
CSCD
关联资源
相关文章
相关作者
相关机构
京公网安备11010802024621