Research On Multi-Qrgan Segmentation Algorithm Of MRI Image Based On Deep Learning

Automatic segmentation algorithms based on MRI images have been widely used in computer-aided diagnosis and treatment.In recent years,the image segmentation technology based on deep learning has developed rapidly,but it still cannot fully fulfill the stringent clinical requirements in terms of accuracy and robustness.For example,the existing segmentation methods are sometimes difficult to distinguish the boundary of the diseased tissue,and the segmentation effect on small organs is not ideal.Aiming at the difficulties in MRI image segmentation,this paper studies the gaps and decoupling problems of network features,and enhances the segmentation accuracy by improving the network structure and loss function.Multi-level feature fusion is often executed in MRI image segmentation networks.However,there are generally differences in semantic and spatial information between multilevel features,which are called feature gaps.When the feature gaps are too large,it will disturb the multi-level feature fusion and affect the segmentation performance.To alleviate this problem,we construct a semantic enhancement module to increase the semantic information of shallow features.A spatial information enhancement module is also constructed to increase the spatial detail information of deep features.These two modules can bridge the feature gaps,achieve robust and efficient multi-level feature fusion,and improve segmentation performance.In addition,in order to bridge multi-level features gaps in the segmentation network,we use the semantic differences between shallow features and deep features at the same spatial position to construct a new semantic loss function to bridge the semantic gaps between shallow features and deep features.We also use the difference in the spatial similarity matrix between deep features and shallow features to construct a new spatial loss function to bridge the gaps in spatial information between deep features and shallow features.These two loss functions can effectively reduce the feature gaps in the segmentation network.We also propose a multi-organ segmentation method on MRI images based on feature decoupling.Different from the existing segmentation methods,we decouple the network features into two parts: low-frequency features and high-frequency features,and propose a new dual-branch feature decoupling network.One branch extracts the features of the smooth region inside the target,called low-frequency feature branch;the other branch extracts the features of the target edge region,called high-frequency feature branch.Through the decoupling method,the features extracted by the two branches are more refined and accurate,thereby effectively improving the segmentation performance.We conducted experiments on two MRI abdominal multi-organ datasets.The results show that our proposed method can effectively improve the accuracy of organ segmentation.This is particularly evident on the edges of organs and small organs.