Research On Multi-Classification Method Of MRI Brain Images Based On Deep Learning

Medical research on human non-brain diseases can be achieved by means of animals,while human brain diseases have many limitations.For example,the human brain and animal brain are essentially different.The human brain can calculate the thinking and memory,and the internal structure is complex and fine.However,direct research on the human brain is contrary to social ethics,so most of the research on the brain in modern medicine is realized by computer-assisted methods.With the rapid development of computer technology,magnetic resonance imaging(MRI)has been widely used in the medical field.It is of practical significance to classify different types of pathological brain images automatically and accurately.This thesis studies the establishment of an automatic detection and classification system for brain images of MRI diseases,which can be classified as health,cerebrovascular disease,neoplastic diseases,degenerative diseases and inflammatory diseases more accurately.The main research works of this thesis are as follows:1.Study the feasibility and significance of auto-encoder(AE)in feature extraction process of MRI brain images,and optimize the auto-encoder.Theis thesis use the improved stacked sparse auto-encoder(SSAE)method to conduct features extracting in MRI brain images.Introducing LI regularization in the AE algorithm,the constraints will be added in the minimized empirical error function.Hence,the model can satisfy the sparsity requirement;2.Aiming at the problem that the training time of deep learning is too long,a mini-batch scale conjugate gradient(MSCG)algorithm is proposed.In the MSCG algorithm,the Levenberg-Marquardt method is used to avoid storing and calculating the second derivative information required by the Newton method.The MSCG algorithm adds scale factors and mini-batch parameters to the conjugate gradient(CG)algorithm.Thereby overcoming the shortcomings of the slow convergence in traditional gradient descent algorithm;3.A hybrid stacked sparse auto-encoder network model is proposed.The SSAE method is used for feature extraction.The softmax classification layer is considered to be stacked with multiple auto-encoder layers to form a hybrid sparse auto-encode classification model.The whole network framework utilize back propagation(BP)algorithm to conduct fine-tuning.This hybrid model can extract features and perform network fine-txme layer by layer,thus improving the classification performance of MRI brain images.