提示
基于眼底彩照多疾病分类的多尺度特征融合网络
首发时间：2025-03-12
摘要：为了提高人工智能辅助诊断多种眼部疾病的效率和准确性，本研究整理了23个公开眼底彩照数据集，归纳出45种眼部疾病作为实验数据，并提出了一种基于眼底彩照多病变分类的多尺度特征融合网络，实现对眼底图像的自动化分类。首先，引入高效代理注意力机制替换基准网络中传统的多头自注意力，减少计算复杂度同时保持优秀的全局上下文建模能力。其次，使用上下文感知注意力机制捕捉局部特征，并使网络对图像的关键病灶区域给予更多关注。最后，在四阶段编码器输出之后，设计了用于增强眼底图像病灶特征的混合多尺度特征增强模块，融合全局和局部信息，捕获多尺度特征表示。所提出的方法加权精确率、加权召回率、加权f1分数、kappa分别达到95.41%、95.18%、95.08%、93.93%。实验结果表明，本研究提出的方法在眼底彩照多分类任务中具有优秀的性能，并优于其他方法。
关键词：
for information in english, please click here
multi-scale feature fusion network for multi-disease classification based on fundus images
abstract：to enhance the efficiency and accuracy of artificial intelligence-assisted diagnosis of various eye diseases, this study compiled 23 publicly available fundus image datasets, encompassing 45 types of eye diseases as experimental data. a multi-scale feature fusion network for multi-disease classification based on fundus images is proposed to achieve automated classification of fundus images. firstly, an efficient agent attention mechanism is introduced to replace the traditional multi-head self-attention in the baseline network, reducing computational complexity while maintaining excellent global context modeling capabilities. secondly, a context-aware attention mechanism is employed to capture local features, enabling the network to focus more on key lesion regions within the images. finally, following the outputs of the four-stage encoder, a hybrid multi-scale feature enhancement module is designed to strengthen lesion features in fundus images, integrating global and local information to capture multi-scale feature representations. the proposed method achieves weighted precision, weighted recall, weighted f1 score, and kappa values of 95.41%, 95.18%, 95.08%, and 93.93%, respectively. experimental results demonstrate that the proposed approach exhibits excellent performance in multi-class classification tasks of fundus images, surpassing other methods.
keywords：
论文图表：
引用
导出参考文献
no.****
动态公开评议
共计0人参与
勘误表
基于眼底彩照多疾病分类的多尺度特征融合网络
评论
全部评论