多尺度特征融合网络的视网膜OCT图像分类

PDF下载HTML

Title:

Retinal optical coherence tomography image classification based on multiscale feature fusion

作者:

韩璐¹, 毕晓君²

1. 哈尔滨工程大学 信息与通信工程学院，黑龙江 哈尔滨 150001;
2. 中央民族大学 信息工程学院，北京 100081

Author(s):

HAN Lu¹, BI Xiaojun²

1. College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;
2. School of Information Engineering, Minzu University of China, Beijing 100081, China

关键词:

视网膜; 光学相干断层扫描; 注意力机制; 空间空洞金字塔; 神经网络; 图像分类; 深度学习; 医学图像

Keywords:

retina; optical coherence tomography; attention mechanism; atrous spatial pyramid pooling; neural network; image classification; deep learning; medical image

分类号:

TP391.7

DOI:

10.11992/tis.202111024

摘要:

目前基于深度学习的视网膜OCT图像分类方法存在网络特征提取能力低、小目标病变分类困难等问题。为此本文提出了一种双分支多尺度特征融合网络，通过加入门控注意力机制，利用深层特征作为选通信号传递给浅层特征，在消除冗余特征的同时，获得更细尺度的抽象信息。同时加入空洞空间金字塔模块，实现在不降低特征图分辨率的同时增大感受野，按不同比例有效捕获全局上下文信息，提高了小目标病变分类精度。实验结果表明，本文提出的方法在视网膜OCT图像分类任务中取得了较好效果，分类准确率达97.9%。

Abstract:

The retinal optical coherence tomography (OCT) image classification method based on deep learning has problems such as low ability of network feature extraction and difficult classification of small target lesions. Therefore, this paper proposes a dual branch multiscale feature fusion network. The gating attention mechanism is added to the vgg16 network, and the deep features are transmitted to the shallow features as gating signals. The redundant features are removed more fine-grained abstract information is obtained. Simultaneously, an atrous spatial pyramid pooling (ASPP) module is introduced to increase the receptive field and capture the global context information in various proportions without reducing the feature map resolution. The ASPP module increases the classification accuracy of small target lesions. The experimental results show that the proposed method has achieved good results in the retinal OCT image classification task, and the classification accuracy has reached 97.9%.