Segmentation And Classification Algorithms Of MR Images For Disease Diagnosis And Treatment

The knee joint is the largest,most complex and strongest sports joint in the human body.The incidence of knee joint injuries is increasing year by year.For common knee diseases,such as anterior cruciate ligament injury and meniscus tear,it is time-consuming and laborious for doctors to analyze MR images manually.There is an urgent need to construct an automated diagnosis method to assist diagnosis.For patients with knee arthritis who need knee arthroplasty,constructing a knee joint model with multimodal information for personalized prosthesis design and preoperative planning can improve the success rate of surgery.At the same time,the thickness of knee cartilage is one of the key indicators for knee arthritis detection.The segmentation of MR images of the knee joint and the construction of a knee joint model fused with multi-modal information play a key role in the diagnosis and treatment of knee arthritis.The main research contents of the paper are as follows:Firstly,a series of pre-processing methods were constructed for clinical MR images to be transformed to the input data of algorithms.By extracting the key information of DICOM data for format conversion and resampling.And performed non-linear filtering and deviation field correction on the images.Experiments proved that standardized data can improve the performance of the algorithm.Then,based on deep learning,an MR analysis method for the differential diagnosis of knee joint diseases was established.Through the analysis of the clinical data in the database,a multi-input convolutional neural network is built.By extracting the largest element of the feature map to solve the problem that the number of MR images is not fixed.Loss function weighting and migration learning were used to solve the problem of sample imbalance and small amount of data,and improved the accuracy of classification.The AUC value can reach 92.6%.Next,an improved MR images segmentation network was built based on U-Net.The accuracy of the segmentation network was improved by residual connection,multi-scale context features extraction module,deep supervision of multi output fusion module and 2.5D input.The experimental results showed that the main evaluation indexes dice similarity coefficient and Hausdorff distance were 93.19%and 5.08 mm respectively.Compared with other existing models,the proposed DRD U-Net achieved better segmentation accuracy.On this basis,the local hospital data set was established.This paper compared and analyzed the local data set and database data,explored the use of generative adversarial learning method to solve the problem of domain migration,and designed an unsupervised segmentation network based on adversarial learning.The main evaluation index dice similarity coefficient of bone region was 90.6%.Finally,in response to the needs of knee arthritis diagnosis and assessment,the segmented MR images were reconstructed in three dimensions,and the thickness of cartilage was quantitatively analyzed,which provided a quantitative means for the grade and progress of knee arthritis.Based on the bone alignment,three-dimensional registration was performed on the knee joint models after three-dimensional reconstruction,which fully integrated the information of multi-modal medical images.For the diagnosis and treatment of diseases,this paper realized the automatic classification and segmentation algorithms of MR images.After establishing the local hospital data set,explored the use of generative adversarial learning to solve the existing domain migration problem.Through registration,a knee joint model fused with multi-modal information was finally established for preoperative planning and personalized prosthesis design.