Application Of Deep Learning Methods In Brain Tumors

Glioma is one of the most common and malignant central brain tumors.Because of its invasive growth causes high mortality and recurrence rates,timely and early diagnosis and treatment are required.As an intuitive and important diagnostic method,gliomas imaging has a guiding role for the follow-up treatments such as surgery and radiotherapy.Therefore,tumor images segmentation and subsequent quantitative analysis are routine and key steps.Magnetic resonance imaging is the main imaging tool for brain tumor currently and provides multi-modal imaging of gliomas.Multi-modal imaging can not only comprehensively analyze anatomical and functional information,but also further study the relationships between modalities,so the tumor structure can be accurately visualized and delineated.However,the amount of 3D tumor MRI data is very large,which is a heavy burden for doctors and experts.If only by manual segmentation it will be very time-consuming,tedious and lack of objectivity.Therefore,automatic segmentation methods are required for accurate,timely,and objective processing.In recent years,deep learning-based segmentation methods have been popular for their wonderful self-learning and generalization capabilities.To capture contextual information and generate semantic features,we propose a novel 3D CNN-based automatic method for brain tumor segmentation.Compared with natural images,medical images are mostly gray scale images with low resolution.Thus this method is based on the encoder-decoder architecture,the main body of which consists of residual modules,encoder adaptation modules,dense fusion modules,and multi-scale optimization modules,and achieves good segmentation results for gliomas.The specific research content is as follows:1.In order to effectively extract representative features,this paper uses an encoder-decoder framework suitable for medical image segmentation.The ideas of widely used residual network structure,dense network structure and Inception network structure are also learned,and 3D modified versions of them are proposed.Then the encoder-decoder structure embedded by residual blocks,dense connections and Inception structure is designed for volumetric medical image segmentation.It can thoroughly explore the advantages of skip connections,residual connections,and densely connections,which used to share information and gradient,so that shallow features and deep features can be combined to achieve finer segmentation.2.In the encoder which is responsible for mining semantic features,we enhance information transmission and deepen the network by effectively stacking residual blocks in the case of medical image resolution adaptation,and progressively extract the high-resolution appearance features and the low-resolution semantic features.The residual connections make the multi-level extraction process easier and more accurate,which lays a good foundation for tumor segmentation.3.In the decoder which undertakes to restore the image resolution,on the one hand,we adopt a dense fusion strategy.In order to reduce the amount of calculation and save memory,the concatenations in dense connections are modified to summations.On the other hand,a multi-scale optimization strategy is used to widen and deepen the network but greatly reduce the network complexity compared to large network structures with equivalent performance.Therefore,this paper uses dense fusion module and multi-scale optimization module to adjust the 3D feature map,thereby reducing the loss of effective information and further improving the segmentation accuracy.4.Ablation experiments were performed to evaluate the performance of key components,including ResNeXt module,dense fusion and multi-scale optimization module.Good results were obtained in multi-modal MRI brain tumor segmentation.In order to better evaluate the model,we chose the classic public data set-BRATS 2015 and compared our proposed model with representative models in BRATS 2015.