Pulmonary DR Image Anomaly Detection Based on Deep Learning

Abstract

The morbidity and mortality in lung cancer is increasing which makes the diagnosis of abnormal lungs particularly important. Because of the advantages in DR image, this paper aimed at two problems in current medical image research: first, it is difficult to completely segment the lung of DR image only used traditional image segmentation methods. This paper replaces the padding in the U-net network model with zero padding to maintain the image size and apply it to the lung DR image segmentation, and finally uses the lung DR image dataset to fine-tuning. Secondly, the results of anomaly detection experiments show that the algorithm would get more complete segmentation of lung DR images. Secondly, because of the insufficient of training set, the idea of multi-classifier fusion is used. Combining Gabor-based SVM classification, 3D convolutional neural network, and transfer learning to achieve a more complete description of features and make full use of the classification advantages of multi-classifiers. The experimental results show that the classification accuracy of this algorithm is 6% higher than that of the Transfer-ImageNet algorithm, 5% higher than SVM, 15% higher than 3D convolutional neural network, and improved 2.5% compared with FT-Transfer-DenseNet3D algorithm.