Cancer Image Recognition Based On Convolutional Neural Network

The dissertation mainly studies cancer image recognition based on convolutional neural network,including classification of breast cancer images and detection of lung nodules.Computer vision technology is used to construct computeraided diagnosis systems for these two types of cancer.Specifically,it includes the following two aspects:(1)Classification of breast cancer histopathological images based on convolutional neural network.Histopathological images of the breast cancer are obtained and fed into a computer-aided diagnosis system,and the system determines whether the image is benign or malignant.For large-scale breast cancer histopathological image dataset Brea KHis,large-scale convolutional neural network is utilized to replace the previous small convolutional network for classification.For solving the problem of training large network,Batch Normalization is introduced to speed up the training of the network.Furthermore,transfer learning and data augmentation are performed to improve the generalization ability.The experimental results show that our model can successfully classify breast cancer images and the accuracy is improved by about 5% compared with the previous networks.(2)Lung nodule detection based on Faster R-CNN.Computed Tomography(CT)images on patient’s chest are fed into the computer-aided diagnosis system to detect the lung nodules.Focusing on the chest CT image dataset LIDC-IDRI,we introduce two-stage object detector Faster R-CNN to detect lung nodules.In order to improve the feature extraction ability of Backbone network,Res Net is adopted instead of VGG.For detecting small lung nodules,Ro IAlign was used instead of Ro IPooling to reduce the error in quantitative operation.The Backbone network was redesigned by dilated convolutions to reduce the information loss of small lung nodules and increase the receptive field.The feature pyramid network is used to fuse the high-resolution feature map in shallow layer and the strong expression competence feature map in deep layer to ensure that small lung nodules can be detected.The experimental results show that the backbone based on dilated convolutions and the feature pyramid network can improve the detection of small lung nodules.