Research On Multi-scale Medical Image Segmentation Based On Data Enhancement And Double Coding

With the popularization of computer and detection technology,medical images have also become an important basis for doctors’ diagnosis.Intelligent medical image processing tools can assist doctors to filter out key information better and faster,and have broad development prospects.Data representation and machine learning are the key technologies for building intelligent medical image processing.The former ensures that the machine learning model has enough information to mine to discover a specific pattern,and the latter ensures that the existing information can be exploited.The appearance of soft tissues in medical images is different.For example,organs will be deformed by factors such as breathing,muscle contraction,and body posture.This phenomenon is called shape multi-scale.In addition,there is a long-tail distribution of medical image data,and most diseases are not common in clinical practice,resulting in small data volume and limited sample distribution.This situation is called multi-scale data volume.Finally,in medical images,there are various forms of disease,and the area scales of different segmentation targets are greatly different.This kind of problem can be called area multi-scale.In this thesis,the multi-scale shape,multi-scale data volume,and multi-scale area problems that exist in medical image segmentation tasks are collectively referred to as multi-scale segmentation problems.From the perspective of expanding data distribution,this thesis solves the problem of multi-scale medical image data shape and data volume by designing a heuristic data augmentation model and an adaptive local elastic deformation search algorithm.At the same time,from the perspective of developing and utilizing existing data,this diser improves the loss function and the encoding method of the model,allows the segmentation model to deeply mine image information,pay attention to small targets and salient targets respectively,and solve the problem of multi-scale area,which provides a basis for follow-up research.A new way of thinking.The main innovations of the dissertation include the following three points:(1)Aiming at the phenomenon of high acquisition cost and multi-scale objects in shapes of medical imaging datasets,a local elastic deformation data enhancement algorithm inspired by soft tissue deformation is designed.According to the three key factors of soft tissue deformation,including local force characteristics,tension constraints and volume invariance constraints,the algorithm first selects a local area,then performs affine on the image in the area,and finally performs elastic deformation simulation.A new kidney imaging tumor dataset is created based on the proposed method to increase the diversity of sample shapes.Experiments show that the proposed algorithm not only works well on medical image datasets,but also performs well on general datasets.(2)Because of the long-tail distribution of medical data,the data volume of tissues or lesions with different appearances varies greatly,resulting in uneven data distribution.However,the generation of elastic deformation is random and cannot balance the data set well.In addition,because there are many local elastic deformation parameters and the location of the deformation is random,more complex parameter tuning is required to find the parameters suitable for the data set.In this dissertation,the adaptive search algorithm is introduced into the local elastic deformation data enhancement,the search space is established first,and then the enhanced sample is automatically selected based on the distance measure,which avoids the complicated parameter selection process,solves the multi-scale problem of data volume,and achieves performance improvement.(3)Aiming at the problem that the existing algorithms are often unable to detect tiny targets and salient targets at the same time and the edge segmentation of complex targets is not clear,that is,the area multi-scale problem,a multiarea multi-scale target loss function and a double-coded segmentation model are proposed.Through The introduction of target area weights and dual-branch encoding enables the model to better focus on tiny targets and extract detailed features on the basis of the original performance.And the practicability of the method is verified on the medical image dataset.