Reasearch On Deep Feature Extraction And Segmentation Of Multimodality MR Images

Multimodality magnetic resonance images have been widely applied in clinical analysis because of its rich feature information.The segmentation of tissues and its lesions based on MR images will provide evidence for diagnosis and treatment.The convolutional neural network method based on U-Net is considered one of the best methods for MR image segmentation.For the different segmentation tasks,the U-Net architecture is improved to enhance its feature extraction ability,which will help improve the segmentation accuracy of the U-Net.The main content of this paper is to propose new depth feature extraction modules,and embed it into the deep convolutional neural network U-Net,so as to achieve the purpose of accurately segmenting the tissues and lesions in multimodality MR images.For MR brain tissues segmentation,a U-Net based multi-neighborhood feature learning segmentation model is proposed.In MR brain,it is difficult to distinguish each tissue(white matter,gray matter,cerebrospinal fluid)in the boundary area.Therefore,in the encoding part of U-Net,the proposed structure with different pooling kernels in parallel is adopted to enlarge the receptive field during down-sampling.In the decoding part,a branch output is added for each up-sampling operation and several groups of channels are generated to collect the multi-neighborhood information in channel dimension.The verification is completed on the official database MRBrain13,the experimental results show that,this method can alleviate the missing of information in the process of encoding\decoding,improve the precision of brain tissues segmentation,especially in the boundary region.For gliomas segmentation,a U-Net based cascaded hybrid network is proposed.In MR glioma,due to the unbalanced distribution of various kinds of glioma samples,U-Net has insufficient discrimination ability for these glioma samples.Aiming at this problem,the generated glioma candidate regions by using 2D-U-Net are sampled to feed into a3D-U-Net for focusing on these samples of false positive samples and wrong segmentation.Finally,in order to take full advantage of 3D multi-view characteristics of glioma,the 3D-U-Net is trained in three directions(axial,sagittal and coronal)and their prediction results are fused.The verification is completed on the official database Bra TS2017,the experimental results show that the new method of U-Net based cascaded hybrid network can enhance the discrimination ability for various kinds of glioma for reducing the false positive and improving the segmentation accuracy.For prostate cancer segmentation,a U-Net based multi-scale spatial attention mechanism segmentation model is proposed.In MR prostate,the target of prostate cancer is difficult to be found due to the small in size and different in shape.Aiming at this problem,an improved multi-scale feature learning technique is used to capture the feature information of different scales of prostate cancers.Additionally,U-Net has rich representation of high-level feature information,multi-scale spatial attention maps are generated by high-level feature maps to better focus on the target.The verification is completed on the official database PROSTATEx,the experimental results show that the multi-scale spatial attention mechanism can enhance the attention of prostate cancer with multi-scale characteristics,and improve the segmentation accuracy of prostate cancer.At present,there is no unified guiding principle for applying U-Net to different MR image segmentation tasks.However,for different segmentation tasks,the U-Net specific feature extraction module proposed above has a certain role in guiding to the segmentation tasks with common characteristics.The multi-neighborhood feature learning method can enhance the feature representation of the target by using the feature information of neighborhood,therefore,it plays a certain role in the segmentation of tissue types across the whole map.The cascaded network can take the region of interest as the center of concern during segmentation,therefore,the method is more effective for the segmentation task with multiple types of samples clustered in a local area.The multi-scale spatial attention mechanism is used to obtain attention of U-Net by adding attention mechanism and multi-scale characteristics of learning,therefore,this method is more effective for small targets with multiple scales.