基于浅层与深层特征融合的胃癌前疾病识别

胃癌前疾病识别对降低癌变风险及胃癌发病率具有重要意义。提出一种基于胃镜图像浅层特征与深层特征融合的胃癌前疾病识别方法。首先,根据胃镜图像性质,手工设计75维浅层特征,包含图像的直方图特征、纹理特征以及高阶特征;然后,基于构建的Resnet、GoogLe Net等卷积神经网络,在其输出层前添加一个全连接层作为图像的深层特征,为保证特征权重一致,全连接层的神经元数目设计为75维;最后,串联图像的浅层与深层特征,使用机器学习分类器,识别胃息肉、胃溃疡和胃糜烂等3类胃癌前疾病。对每种疾病收集了380张图像,并以4:1的比例划分为训练集和测试集,然后基于该数据集,分别采用传统机器学习、深度学习、特征融合等3种方法进行模型训练和测试。模型在测试集上的结果显示,所提出的特征融合方法识别准确率高达95.18%,优于传统的机器学习方法(74.12%)和深度学习方法(92.54%)。所提出的方法能够充分利用浅层特征与深层特征,为医生提供临床决策支持以辅助胃癌前疾病诊断。

Precancerous Diseases Classification Based on Fusion of Shallow and Deep Features

Precancerous disease recognition is of great significance in reducing the risk of gastric cancer. This paper proposed a method for identifying precancerous diseases based on the fusion of shallow and deep features of gastroscopic images. Firstly,according to properties of gastric images,75-dimensional shallow features were designed manually,including histogram features,texture features and higher order features. Secondly,based on the networks of Resnet and GoogLeNet,we added a fully connected layer before the output layer to extract the deep features of the images. To ensure consistent feature weights,the dimension of the added fully connected layer was 75. Finally,the shallow features were merged with deep features. Machine learning classifiers were used to identify three types of precancerous diseases,namely gastric polyps,gastric ulcers and gastric erosions. We collected 380 images for each disease,and 75% were used as training sets,the others were used as testing sets. We conducted experiments using traditional machine learning,deep learning and feature fusion proposed in this paper. Experiment results showed that the recognition accuracy of the feature fusion method proposed was as high as 95.18%,significantly better than that of traditional machine learning methods (74.12%) and deep learning methods (92.54%). This proposed method made full use of the shallow features and deep features to provide clinical decision support for doctors and assist in the diagnosis of precancerous diseases.