Research On Image Classification And Segmentation Of Bladder Cancer Based On Deep Learning

Bladder cancer is a type of malignant tumor with high morbidity and mortality.It’s a common disease in male urinary system.Magnetic resonance imaging(MRI)can produce tissue cross-sectional images from different orientations with different tissue characteristics.,which is an important method in medical diagnosis.MRI can realize high-resolution imaging of soft tissues and is widely used in cancer diagnosis of abdominal organs.Magnetic resonance examination usually collects images from multiple sequences,and each sequence contains a large number of images.In tumor detection and diagnosis doctors need to screen out a large number of magnetic resonance images.In the diagnosis of bladder cancer,quick and accurate segmentation of bladder wall and tumor area from MR images is of great significance for computer-aided diagnosis.At present,the tumor classification and segmentation from magnetic resonance images are mainly performed manually,which is a tedious and heavy workload.The accuracy of classification and segmentation is limited by the experience of doctors.Automated classification and segmentation of medical images has an unmet clinical need.In recent years,deep learning technology has been widely used in the field of computer vision.Convolutional neural networks have achieved satisfactory results in the classification and segmentation from medical images.This thesis applies deep learning methods to MRI for automatic classification and segmentation of bladder cancer through a convolutional neural network.The main work and contributions of this study are as follows:Inspired by the idea of transfer learning,a classification network of MRI images of bladder cancer based on Inception-ResNet was proposed.This network combines the advantages of multi-scale receptive fields and residual connections brought by convolution kernels of different sizes in Inception-ResNet,which can effectively extract deep-level features.In order to extract the shallow features of tumors in magnetic resonance images,based on Inception-ResNet,a shallow feature extraction network branch is constructed in this thesis.This branch includes three layers of convolutions,and batch normalization is performed after each convolution.Furthermore,a Dropout operation is performed,and the features extracted from the two parts are fused and then classified.In this thesis,the labeling and data preprocessing were performed to construct a magnetic resonance data set for bladder cancer,and then the network was trained.This thesis is compared with Inception-V3 and VGG-16 networks,respectively.Experimental data show that the proposed network has higher classification accuracy.Based on the classification of MRI images of bladder cancer,an automatic segmentation network of MR images of bladder cancer based on U-Net and attention mechanism was proposed.The boundaries of bladder cancer on MR images could be complex and fuzzy,and high-resolution information is required for accurate segmentation.U-Net’s encoder-decoder"U"-shaped structure performs target recognition based on low-resolution information and achieves accurate segmentation and positioning based on high-resolution information.Therefore,U-Net is suitable for medical image segmentation tasks.In order to expand the receptive field to achieve long-range dependence in bladder cancer images and make full use of the global information of the image,considering that the U-Net network is shallow and not easy to deepen,this thesis introduces an attention mechanism for improving U-Net network.This can be more flexible by capturing long-range dependencies without deepening the depth of the network,while maintaining the same input scale and output scale.After training on the bladder cancer MRI dataset,the experimental test results show that the U-Net network based on the attention mechanism performs better than original U-Net network in bladder cancer segmentation.