Research On Parallelization Of Preprocessing For Automatic Detection Of Pulmonary Nodules In 3D CT Images

As one of the most morbid cancers in the world,lung cancer has far more mortality than other cancers and has become one of the most important threats to human health Early detection and treatment of lung cancer can greatly improve patients’ survival rate.Pulmonary nodules,a common lesion in the lungs,usually appear as an early manifestation of lung cancer.Therefore,early detection of pulmonary nodules plays a key role in the treatment of lung cancer.In order to help outpatients reduce workload and improve diagnostic accuracy,computer-aided diagnosis(CAD)systems have been introduced for therapeutic institutions.In order to reduce the overall computational time of a CAD system and improve its practicability,this thesis conducts parallel research on the pre-processing steps in the automatic detection system of pulmonary nodules.This thesis works on parallelization of the automatic detection system of lung nodules from two dimensions,data and operation.We first propose a parallelization architecture based on static data segmentation.The 3D lung CT images are grouped in the form of image slices,and the grouped data is parallelized.Then,a pipeline-based parallelization architecture is proposed,and mutually independent steps are distributed to the respective processing units,and the pre-processing operations are performed in a parallel pipeline manner.Finally,a bus-based parallelization architecture is proposed.Based on the pipeline,a control unit is added to dynamically assign computing tasks to the various execution units.At the end of this thesis,the experimental performance of three architectures is compared.The experimental data is from the LUNA16 public database,and we take the pre-processing steps of extracting the foreground,segmenting the lung parenchyma,repairing the lung parenchyma,and extracting the candidate nodules.The experiment consists of two experiment groups,one is based on the message queues in a local environment,and the other is in a simulated network environment based on Docker.The experimental results show that in the local environment,static data grouping and busbased parallelization architecture have achieved good results,and the computation power utilization rate is more than 80%.The pipeline-based parallelization architecture is not suitable for pre-processing of automatic pulmonary nodule detection systems.In the network environment,the static grouping-based parallelization architecture outperforms the others.The pipeline-based parallelization architecture performance is least affected by the network transmission efficiency.The bus-based parallelization is the most severely degraded in the network environment.In the end,we conclude that in most cases,the parallelization architecture based on static data grouping can effectively speed up the system;Bus-based parallelism architecture can achieve better results in the case of uneven data distribution and instant communication speed between execution units.