融合PVTv2和动态感知的视网膜分级算法

doi:10.19665/j.issn1001-2400.20240801

摘要/Abstract

摘要：

针对视网膜病变分级中存在的图像质量差异和病变区域特征识别困难等问题,提出一种融合PVTv2和动态感知的视网膜分级算法,即先对Eye-Quality数据集进行质量评估,随后进行病变分级。该算法首先对数据集进行灰度化、高斯滤波等操作,凸显病灶区域与噪声背景的差异,其后由PVTv2骨干网络对视网膜图像进行多尺度特征提取,实现多尺度特征纹理的全面捕捉;然后利用并行化区域感知模块和通道重建单元模块,抑制背景区域的干扰信息并聚焦病变特征区域,提升模型的特征识别能力;再通过动态自适应特征融合模块,建立全局特征与局部特征间的动态联系,获取深层语义信息与边缘细节信息;最后采用混合损失函数缓解样本分布不均匀问题,进一步增强模型的视网膜分级效果。质量分级实验中在Eye-Quality数据集上准确率和精确度分别为88.68%和87.72%。在此基础上进行病变分级,使用准确率为评价指标,其优质、可用和拒绝分别为80.23%、74.37%和73.73%。结果表明图像质量的差异将影响病变分级效果,为视网膜分级智能辅助诊断提供了新窗口。

关键词: 病变分级, 质量分级, 并行模块, 自适应模块

Abstract:

To address the challenges of image quality variations and difficulty in lesion area recognition in retinal lesion grading,a novel retinal grading algorithm that integrates PVTv2 and dynamic perception is proposed.Initially,the Eye-Quality dataset undergoes the quality assessment followed by lesion grading.The algorithm first preprocesses the dataset with grayscale conversion and Gaussian filtering to enhance the contrast between lesion regions and background noise.Subsequently,the PVTv2 backbone network performs multi-scale feature extraction on retinal images,achieving the comprehensive capture of multi-scale textural features.Parallel region perception modules and channel reconstruction units are employed to suppress background interference and focus on lesion features,enhancing feature recognition capabilities.A dynamic adaptive feature fusion module establishes connections between global semantic information and edge details.Finally,a hybrid loss function alleviates class imbalance issues,further improving the retinal grading performance.In quality grading experiments on the Eye-Quality dataset,the accuracy and precision are 88.68% and 87.72%,respectively.Then,using accuracy as the evaluation metric for lesion grading based on this foundation,the rates for excellent,usable,and rejectable categories are 80.23%,74.37%,and 73.73%,respectively.These results highlight the impact of image quality differences on lesion grading effectiveness,providing a new avenue for intelligent auxiliary diagnosis in retinal grading.

Key words: lesion grading, quality grading, parallel module, adaptive module

中图分类号:

TP391

梁礼明, 金家新, 李俞霖, 董信. 融合PVTv2和动态感知的视网膜分级算法[J]. 西安电子科技大学学报, 2024, 51(6): 159-170.

LIANG Liming, JIN Jiaxin, LI Yulin, DONG Xin. Retina grading algorithm integrating PVTv2 and dynamic perception[J]. Journal of Xidian University, 2024, 51(6): 159-170.

图/表 15

图1

图2

图3

图4

图5

表1

表2

图6

表3

图7

表4

表5

图8

表6

图9

参考文献 26

[1]	梁礼明, 董信, 雷坤, 等. 融合注意力谱非局部块的视网膜图像质量分级[J]. 西安电子大学学报, 2024, 51(4):102-113.
	LIANG Liming, DONG Xin, LEI Kun, et al. Retinal Image Quality Grading by Fusion of Non-Local Blocks of Attention Spectrum[J]. Journal of Xidian University, 2024, 51(4):102-113.
[2]	韩冰, 高路, 高新波, 等. 边界加权的甲状腺癌病理图像细胞核分割方法[J]. 西安电子科技大学学报, 2023, 50(5):75-86.
	HAN Bing, GAO Lu, GAO Xinbo, et al. Boundary-Weighted Nuclei Segmentation Method in Thyroid Cancer Pathological Images[J]. Journal of Xidian University, 2023, 50(5):75-86.
[3]	GUO T, LIANG Z, GU Y, et al. Learning for Retinal Image Quality Assessment with Label Regularization[J]. Computer Methods and Programs in Biomedicine, 2023, 228:107238.
[4]	SHEN Y, SHENG B, FANG R, et al. Domain-Invariant Interpretable Fundus Image Quality Assessment[J]. Medical Image Analysis, 2020, 61:101654.
[5]	ZHANG P, HU T, HAN X, et al. Image Quality Assessment of Diabetic Retinopathy Based on ADD-Net[J]. IEEE Access, 2023, 11:105130-105139.
[6]	FU H, WANG B, SHEN J, et al. Evaluation of Retinal Image Quality Assessment Networks in Different Color-Spaces[C]// Medical Image Computing and Computer Assisted Intervention. Berlin:Springer, 2019:48-56.
[7]	WANG W, XIE E, LI X, et al. PVTv2:Improved Baselines with Pyramid Vision Transforme[J]. Computational Visual Media, 2022, 8(3):415-424.
[8]	XU S, ZHENG S C, XU W, et al. HCF-Net:Hierarchical Context Fusion Network for Infrared Small Object Detection(2024)[J/OL]. [2024-03-16]. https://arxiv.org/abs/2403.10778.
[9]	VASWANIA, SHAZEER N, PARMAR N, et al. Attention is All You Need[C]// 31st Conference on Neural Information Processing Systems(NIPS 2017). San Diego: NIPS, 2017:1-11.
[10]	SHIB, GAI S, Darrell T, et al. Refocusing is Key to Transfer Learning(2023)[J/OL]. [2023-01-01]. https://arxiv.org/abs/2305.15542.
[11]	MOU L, ZHAO Y, CHEN L, et al. CS-Net:Channel and Spatial Attention Network for Curvilinear Structure Segmentation[C]// Medical Image Computing and Computer Assisted Intervention-MICCAI 2019. Berlin:Springer, 2019:721-730.
[12]	WANG Q, WU B, ZHU P, et al. ECA-Net:Efficient Channel Attention for Deep Convolutional Neural Networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2020:11534-11542.
[13]	LI J, WEN Y, HE L. Scconv:Spatial and Channel Reconstruction Convolution for Feature Redundancy[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2023:6153-6162.
[14]	LIX, WANG W, HU X, et al. Selective Kernel Networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2019:510-519.
[15]	YANG J, QIU P, ZHANG Y, et al. D-Net:Dynamic Large Kernel with Dynamic Feature Fusion for Volumetric Medical Image Segmentation(2024)[J/OL]. [2024-10-17]. https://arxiv.org/abs/2403.10674.
[16]	HOU Q, ZHOU D, FENG J. Coordinate Attention for Efficient Mobile Network Design[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2021:13713-13722.
[17]	MUKHOTI J, KULHARIA V, SANYAL A, et al. Calibrating Deep Neural Networks Using Focal Loss[J]. Advances in Neural Information Processing Systems, 2020, 33:15288-15299.
[18]	REZAEI-DASTJERDEHEI M R, MIJANI A, FATEMIZADEHE. Addressing Imbalance in Multi-Label Classification Using Weighted Cross Entropy Loss Function[C]// 2020 27th National and 5th International Iranian Conference on Biomedical Engineering(ICBME). Piscataway:IEEE, 2020:333-338.
[19]	CUADROS J, BRESNICK G. EyePACS:An Adaptable Telemedicine System for Diabetic Retinopathy Screening[J]. Journal of Diabetes Science and Technology, 2009, 3(3):509-516.
[20]	郑雯, 沈琪浩, 任佳. 基于Improved DR-Net算法的糖尿病视网膜病变识别与分级[J]. 光学学报, 2021, 41(22):72-83.
	ZHENG Wen, SHEN Qihao, REN Jia. Identification and Grading of Diabetic Retinopathy Based on the Improved DR-Net Algorithm[J]. Acta Optica Sinica, 2021, 41(22):72-83.
[21]	OU F Z, WANG Y G, ZHU G. A Novel Blind Image Quality Assessment Method Based on Refined Natural Scene Statistics[C]// 2019 IEEE International Conference on Image Processing(ICIP). Piscataway:IEEE, 2019:1004-1008.
[22]	YAN Q, GONG D, ZHANG Y. Two-Stream Convolutional Networks for Blind Image Quality Assessment[J]. IEEE Transactions on Image Processing, 2018, 28(5):2200-2211.
[23]	SONG J W, YANG R Y. Feature Boosting,Suppression,and Diversification for Fine-Grained Visual Classification[C]// 2021 International Joint Conference on Neural Networks(IJCNN). Piscataway:IEEE, 2021:1-8.
[24]	梁礼明, 雷坤, 詹涛, 等. 基于锐度感知最小化与多色域双级融合的视网膜图片质量分级[J]. 科学技术与工程, 2022, 22(32):14289-14297.
	LIANG Liming, LEI Kun, ZHAN Tao, et al. Retinal Image Quality Based on Sharpness Perception Minimization and Multi-Gamut Bi-Level Fusion[J]. Science Technology and Engineering, 2022, 22(32):14289-14297.
[25]	LIU D, ZHAO L, WANG Y, et al. Learn from Each Other to Classify Better:Cross-Layer Mutual Attention Learning for Fine-Grained Visual Classification[J]. Pattern Recognition, 2023, 140:109550.
[26]	KARKERA T, ADAK C, CHATTOPADHYAY S, et al. Detecting Severity of Diabetic Retinopathy from Fundus Images:A Transformer Network-based Review(2023)[J/OL]. [2023-01-03]. https://arxiv.org/abs/2301.00973.

	训练集
	DR:0	DR:1	DR:2	DR:3	DR:4	All
优质	6 342	699	1 100	167	39	8 347
可用	1 353	103	283	79	58	1 876
拒绝	1 544	109	426	87	154	2 320
整体	9 239	911	1 809	333	251	12 543

	测试集
	DR:0	DR:1	DR:2	DR:3	DR:4	All
优质	5 966	886	1 354	199	65	8 470
可用	3 201	359	721	145	133	4 559
拒绝	2 195	153	569	104	199	3 220
整体	11 362	1 389	2 644	448	397	16 249

Model	Acc/%	Precision/%	Recall/%	F1/%
M1	87.94	87.63	84.81	86.05
M2	88.34	86.96	86.69	86.82
M3	88.36	87.88	85.62	86.61
M4	87.65	87.07	84.74	85.70
M5	88.58	87.51	86.45	86.85
M6	88.68	87.72	86.48	86.95

Methods	Acc/%	Precision/%	Recall/%	F1/%
NBIQA^[21]	79.17	76.41	75.09	74.41
TS-CNN^[22]	79.26	79.76	74.46	74.81
ResNet-50+FDM^[23]	87.23	86.40	84.54	85.13
ResNeSt^[24]	87.35	85.87	85.07	85.44
ResNet50^[1]	88.59	87.56	86.10	86.74
文中算法	88.68	87.72	86.48	86.95

Methods	Acc/%	AUC/%
未进行质量分级	79.25	84.04
进行质量分级	80.23	83.72