Colorectal Cancer Segementation From CT Images Based On U-Net

Colorectal cancer ranks the third and second in the incidence and mortality of malignant tumors globally,and fifth in the incidence and mortality of malignant tumors in China.CT images are widely used in clinical practice for diagnosing colorectal cancer,and tumor segmentation is an important first step in tumor analysis.However,most of the tasks of segmenting lesions are done manually by doctors,which is tedious,laborious,low consistency and poor reproducibility.Therefore,accurate automatic segmenting methods need to be developed urgently.In this study,we firstly analyze the characteristics of rectal tumors,such as large differences in shape and size,wide distribution range in abdomen and large original images.And then we investigate problems of using the popular u-net and its variants to segment colorectal tumors: insufficient contextual information of small volume,low efficiency,out of memory for big volume.In order to solve these problems,we proposed a model which is friendly to GPU memory by combining the existing methods.Firstly,we adopted the framework of Roi aware u-net model,which locates the Roi position through positioning module,and then partially decodes the features,which can greatly reduce the GPU memory.Since the intermediate variable of the reversible residual module can be obtained by inference for backward,it is unnecessary to preserve the intermediate variable and the module could significantly reduce the GPU memory.Therefore,we use the reversible residual module to replace the residual module of the original model.Experiments show that under the premise of similar performance with original residual modules,reversible residual modules occupied less GPU memory when training.Because existing method based on U-Net directly encode the original resolution images,which would cause high memory consumption,we proposed a down-sampling module for multi-direction,the module is divided into three branches,each branch corresponds to a direction without sampling,and the another two directions were down-sampled.The module each branch retains the one part of original resolution information.Experimental results show that although the proposed module has a performance loss of about 1% compared with the original resolution encoding method,but can significantly reduce the GPU consumption.In addition,in order to enhance the feature grasping ability of the model,we embedded an attention module in the decoding part.Since the decoding stage is minor had limited GPU memory consumption,and the adding attention module only occupied minor GPU memory.Experimental results show that the attention module could enhance the segmentation performance of the model.In order to show the advantages of proposed model in comparison experiment,the model in this paper used large image blocks for training,the results show that the model can significantly reduce the false positive results,and improve the accuracy.