One-pass Multi-task Network For Brain Tumor Segmentation

Glioma is a common malignant brain tumor with highest mortality and prevalence,which seriously endangers human life and health.With the rapid development of modern imaging technology,especially magnetic resonance imaging(MRI),doctors can diagnose and evaluate brain tumors based on multi-modal MRI images,so as to develop effective treatment methods.Therefore,timely and accurate segmentation of brain tumors is very important for doctors in many processes,such as making treatment plans for patients,performing surgery and prognosis follow-up.However,the manual segmentation of brain tumors is a laborious and time-consuming task,and is easily affected by subjective factors.Therefore,researchers are devoted to the development of automatic brain tumor segmentation technology.However,because the shape of brain tumor is changeable,its structure is complex and its intensity is not uniform,and there is a serious class imbalance problem in brain tumors segmentation,the research of automatic brain tumor segmentation methods is a very challenging task.With the vigorous development of deep learning in recent years,more and more researchers have applied it to the field of brain tumor segmentation,and achieved good results.In particular,model cascade strategy has recently been widely used in medical image segmentation tasks.By running a series of independent deep models,this strategy effectively alleviates the common class imbalance problem,achieves coarse-to-fine segmentation,and obtains promising segmentation performance.Unfortunately,we observe that this strategy has high system complexity and neglects the correlation between deep models.Therefore,in order to solve the above shortcomings,we propose a one-pass multi-task convolution neural network.Firstly,we integrate multiple segmentation tasks in model cascade into a deep model,which is One-pass Multi-task Network(OM-Net).It can not only utilize the correlation between tasks to save a lot of training parameters,but also needs only one-pass calculation and obtains excellent segmentation results of brain tumors,which solves the problem of class imbalance better than model cascade strategy.In addition,based on the special structure of OM-Net,we design an online data transfer strategy,which can make certain tasks get more training data and improve data utilization.Furthermore,we propose a training strategy based on curriculum learning,which introduces tasks to OM-Net in an order of increasing difficulty level,so as to train OM-Net more effectively.Secondly,we further explore the correlation by sharing the prediction results between tasks,and propose the Cross-task Guided Attention(CGA)module.With the guidance of the coarse segmentation results generated by the previous task,CGA can adaptively recalibrate the channel-wise feature responses according to the statistical information of specific categories.Compared with popular SE(Squeeze & Excitation)blocks,CGA can utilize the correlation between tasks to provide additional guidance information to help learn category-specific channel attention.Finally,we propose a simple and effective post-processing method to refine the segmentation results of OM-Net,in order to further improve the segmentation performance.Systematic experiments are performed on several datasets to justify the effectiveness of the proposed methods.The proposed methods achieve state-of-the-art performance on BraTS 2015 and BraTS 2017 datasets respectively.With the proposed techniques,we also won the third place in BraTS 2018 challenge among 64 participating teams.