An Image Segmentation Algorithm For Intracranial Hemorrhage Based On Semi-supervised Mechanism

Intracranial hemorrhage(ICH)is a serious disease with high incidence,high recurrence rate and high disability rate.Deep learning has been widely used in medical image segmentation tasks.The medical image segmentation model based on deep learning can automatically learn comprehensive and high-level features from training images,and then get more accurate segmentation results.Generally speaking,the performance of deep learning model is closely related to the feature extraction ability of the model and the number of training data sets.In terms of enhancing the feature extraction ability of deep learning models,attention mechanisms have been proven to be an effective and highly interpretable method.However,there are still some issues with the existing attention mechanisms used in computer vision analysis,including:(1)the mechanism of directly using pooling operations to condense features when calculating channel attention weights is too simple,which can easily ignore important information and affect the calculation of attention weights.(2)Generating only one spatial attention weight map when calculating spatial attention weights is difficult to generalize the weights of the same position on all channels,and the accuracy of generating spatial attention maps using only convolution operations is limited.In terms of improving the quality of datasets,most segmentation algorithms currently rely on pixel-level annotations,which are often expensive,tedious,and labor-intensive.Compared with other conventional computer vision fields,the scarcity of high-quality annotated datasets is particularly severe in medical image data due to privacy issues and the need for experienced clinical doctors to perform annotations.In recent years,semisupervised medical image segmentation has been proven to be an effective method.However,current semi-supervised medical image segmentation methods are mostly focused on the research of one of the strategies,and there are few studies that combine multiple strategies.Using only one of these strategies cannot fully utilize the information contained in the unlabeled data.Regarding the above two issues,this paper proposes the following methods to address them by improving the feature extraction ability of network models and the information mining capability of datasets:1.This paper proposes a novel attention mechanism-fine-grained global attention mechanism.The proposed method consists of two parts: fine-grained channel attention mechanism and fine-grained spatial attention mechanism.The fine-grained channel attention mechanism improves the traditional channel attention calculation method by introducing dilated convolution and depth-wise convolution in the feature compression stage,which increases the calculation accuracy of channel attention with only a small amount of additional computational cost.The finegrained spatial attention mechanism solves the problem that a single spatial attention map is difficult to summarize the weights of the same position on all channels by grouping the feature maps,and calculates attention weights separately in the W and H directions to obtain more precise attention weights.Finally,through extensive experiments,the effectiveness and versatility of the proposed attention module are validated on different types of medical image datasets.2.This paper proposes a novel semi-supervised segmentation model for intracranial hemorrhage CT images.We effectively integrate semi-supervised learning based on pseudo-label co-training and consistency regularization to fully utilize the potential information of unlabeled images.In addition,we introduce CNN and Transformer dual architectures into the backbone network for co-learning,using different network architectures to generate pseudo-labels for semi-supervised training,so that the two architectures can learn from each other’s obtained information.Finally,we conduct extensive experiments on intracranial hemorrhage CT dataset to demonstrate the effectiveness and superiority of the proposed architecture.