Research On Weakly Supervised Segmentation Method For MRI Glioma Images

Glioma is the most common primary intracranial tumor.MRI is the preferred method for glioma screening.Multimodal MRI images can help doctors obtain detailed information about gliomas to analyze and treat glioma.However,it is difficult for doctors to accurately identify gliomas.Therefore,using image segmentation technology to realize the visualization of glioma lesions can greatly improve the diagnostic accuracy of gliomas.At the same time,the segmentation results can also provide reference data for surgical evaluation.However,in the case of medical images,due to the large differences between the image generation hardware and the shooting organs themselves,using a fully supervised training method to segment glioma requires manual pixel-level labeling of the training data set.It will consume a lot of human and material resources and time.Using weak supervision to train glioma images can make up for the disadvantage of fully supervised learning that requires a large number of pixel-level labels.Class Activation Map(CAM)is an important technique in weakly supervised segmentation,which enables image segmentation without pixel-level labels.However,for MRI glioma images,CAM cannot completely cover glioma.In order to solve this problem,two different methods are proposed to further improve the performance of CAM.The specific work is as follows:(1)A global weighted average pooling network that fuses the grayscale information of MRI images is proposed.Due to the global average pooling(GAP),the convolutional network treats important regions and non-important regions in the feature map equally during training.In this paper,by introducing the weight matrix before GAP,different weights are learned for different positions of the feature map,which can effectively solve the problem that the network treats all regions of the feature map equally.At the same time,we take advantage of the unique advantages of MRI images and use the grayscale difference between the glioma area and the non-glioma regions in glioma images fuse the low-level grayscale information from medical images with high-level semantic information extracted from the network.Give full play to the advantages of different levels of topographic maps and learn the weight of different positions in topographic maps together.Our experimental results on the popular Bra TS2019 medical image dataset show that the proposed method can very well improve CAM performance and aid CAM in matching object boundaries.Meanwhile,in the DSC evaluation,the proposed method achieves a score of 64.1%,an improvement of 4.6% over the recent research method.(2)A membership matrix supervised weak learning strategy based on multi-level subdivision classification is proposed.First,this paper proposes to use a multi-level fine-grained classification strategy to classify the dataset in more detail.This classification strategy can enable the convolutional network to learn more detailed features of the input data,thereby enhancing the CAM.Secondly,it introduces the idea of fuzzy clustering,uses CAM to obtain the class centers of target and non-target,and infers the membership degree of each pixel in the input data for different categories,and then obtains the membership degree matrix.Finally,use the membership matrix and CAM to construct the loss to supervise the training of the network to help the convolutional network mine the target boundary information,so that CAM can better fit the target boundary.A large number of experiments at BRATS 2019 demonstrate that the approach presented here can indeed improve CAM performance.The proposed method is 17.1% higher than the baseline method.Compared with the recent research,it increase by 9%.