Research On Brain MRI Unsupervised Anomaly Detection Algorithm Based On Convolutional Autoencoder

Brain is the key part of the human body and the most advanced part of the central nervous system.Anomaly or diseases in the brain can make a major impact on the basic survival ability of human and even lead to death.Lesions in the brain or spinal cord tissue are called brain tumors.The five-year survival rate of patients is only about 35%.But the boundaries of the brain tissue are blurred so that the lesions(or anomaly)and normal tissue are difficult to define.The automatically detection and location of brain lesions are the prerequisites for further treatment and diagnosis.However,doctors not only need rich medical experience and professional knowledge to complete these tasks,but also cost a lot of time and energy.Although there have been some researches devoted to the automated auxiliary diagnosis of the brain(such as brain tumor segmentation,brain tumor classification,etc.),these methods rely heavily on a large amount of data while medical data is scarce,unbalanced,diverse and has extremely high cost of annotation that is different from traditional data.This problem makes data collection and annotation become a major bottleneck restricting the development of automated medical assisted diagnosis.In order to reduce the dependence on data and its annotations,and at the same time make full use of the characteristics of medical data that is easy access to the data of healthy human,this paper proposes a brain MRI unsupervised anomaly detection algorithm based on convolutional autoencoders.This paper mainly uses the convolutional autoencoder as the base model,and uses unsupervised learning to realize the detection and localization of brain MRI lesions.Specifically,the algorithm has only seen normal brain MRI images during the training stage,but it can distinguish between normal and abnormal brain MRI images during the testing stage and has a certain ability to locate the lesions.The algorithm is based on the idea of reconstruction.The main assumption is that the algorithm completes the modeling of the normal brain MRI data distribution during the training process.Since the algorithm only learns to reconstruct the normal brain MRI,it can only reconstruct the normal brain well and fail to reconstruct the abnormal brain in the testing phase.The lesion detection is realized by the residual of the reconstructed image and the original image.The main work of this paper is as follows:1)According to the setting of unsupervised anomaly detection,the brain MRI image unsupervised anomaly detection dataset Brain AD is collected and constructed to complete the analysis and preprocessing of the brain MRI image;2)Research on brain MRI image anomaly detection algorithm based on the convolutional autoencoder,and explore the application of reconstruction-based abnormality detection algorithm in brain MRI image;3)Research on brain MRI image anomaly detection algorithm based on the memoryaugmented convolutional autoencoder,and explore the effectiveness of memory network on the extraction of prototypical features of normal samples and its impact on the performance of brain MRI image anomaly detection;4)Research on the brain MRI image anomaly detection algorithm based on the memory-augmented convolutional autoencoder with adversarial learning,to explore the effectiveness of the adversarial learning on capturing data distribution and the impact of low-dimensional adversarial learning representation on the performance of brain MRI image anomaly detection.The extensive experiments show that the brain MRI unsupervised anomaly detection algorithm based on convolutional autoencoders can still effectively distinguish between normal brain and abnormal brain MRI images without using abnormal brain MRI images to participate in training.