| Glioma is the tumor with the largest proportion and the highest probability of malignancy among brain tumors.Magnetic resonance imaging(MRI)has a high imaging resolution for soft tissues and has been widely used in the field of brain disease diagnosis.Accurate and rapid segmentation of the tumor area in the glioma magnetic resonance image can obtain physiological information such as the location and size of the tumor,and then assist clinical experts in establishing a diagnosis and treatment plan for the patient.This work mainly carried out the following three aspects of research:(1)In order to meet the high-precision segmentation requirements of glioma images,the glioma segmentation problem was transformed into a glioma feature extraction task.However,the texture features and edge features in glioma images are relatively complicated,and Deep Labv3+ has the problem of inaccurate feature extraction.For this reason,this work added a convolutional block attention module to its backbone network,and assigned different weights to it according to the importance of different feature information.This work used a smoother Mish activation function to maintain high accuracy in the segmentation of glioma lesions.At the same time,it can avoid gradient explosion and overfitting in the training process.The Kappa coefficient of the final model for the Bra TS2019 dateset reached 90.74%,and the Dice coefficient reached 91.11%.Compared with the original basic model,the performance of this model improved by 0.95% and 0.91%,which is close to the highest level of similar researches.It shows a better adaptability and showing better adaptability and more efficient learning ability of glioma characteristics.(2)In order to meet the needs of lightweight segmentation of glioma images,a glioma magnetic resonance image segmentation model based on Transformer network was proposed.Aiming at the computational load problem of the self-attention module,the axial attention module was introduced to reduce the amount of calculation.The global-local training strategy was used to complete the simultaneous extraction of axial large-scale features and local texture features,thereby the glioma can be segmented.The average pixel accuracy of the final model for the Bra TS2019 dateset is 93.91%,which is better than other image segmentation models,and the average intersection ratio is 71.51%.Through the comparison of the calculation parameters of the models,the lower dependence of this model on the computing power was verified,and it is suitable for the mobile terminal to process the low computing power scene of the glioma images.(3)In order to reduce the workload of clinical experts in reading images,a similarity feature learning model for glioma magnetic resonance images based on the VGG twin network was proposed.According to the similarity threshold setting between the core frame image and the discriminated image,the compression ratio of the image frame number was controlled to realize the intelligent capture of the key frame image.In the comparison experiment with the manual extraction of key frames by clinical experts,the average accuracy of the model reached 92.97%.The application software was developed based on the algorithm of this work,and the clinical test was carried out in Grade A tertiary hospital.Integrating clinical case MRI images and diagnosis report results,the feasibility and clinical value of the proposed algorithm were preliminarily verified.Quantitative evaluation and analysis of segmentation results were completed through full communication with clinical experts,and the improvement scheme of the algorithm was also discussed. |
PDF Full Text Request