Research On Frequent Approximate Subgraph Mining And Classification Methods Targeting At Brain Medical Images

Brain diseases have been the second cause of global human death since 2002.Brain medical images have been an important tool for doctors' diagnosis for the explicitly visualizing ability to normal and abnormal brain structures in a non-invasive manner.This leads to a large number of brain medical images generated from hospitals.The rapidly growing of images has made medical image-based research transform from a hypothesis-driven to a data-driven research manner.This also facilitates computer-aided analysis aiming at brain medical images to be an important topic in the field of healthcare big data.In the domain of brain medical image analysis,how to effectively extract features from images has always been a hotspot topic.Image features mainly fall into two categories: handcrafted and automatically learning features.Hand-crafted features consist of low-level visual features and high-level semantic features.Objects extracted from images and their spatial topological relations are the high-level semantic features of images and the commonly used representation method is graph models.However,existing graph models aiming at medical images do not consider the symmetry and connection between left and right hemispheres.Moreover,important pattern mining aiming at brain medical images has been well investigated.Corners can describe the morphology and grayscale changes of images with less information.They are effective low-level visual features of medical images.To date,brain medical image analysis based on corners generally aims at image registration.How to utilize corner to classify brain medical images has not been effectively investigated.The above two topics are brain medical image analysis based on hand-crafted features.Deep learning based image analysis methods directly take images as input to automatically learn features.They do not require any pre-defined feature.Among deep learning models,convolutional neural networks(CNNs)can better recognize general images(e.g.,scenery)than human beings.Currently,some CNNs have developed on brain medical images.However,they have the following shortages:(1)Most of the methods utilize data which needs a series of complex preprocessing steps but not brain medical images as the direct input of convolutional neural network.(2)These methods aim at cross-sectional images.Longitudinal medical images can display pathological changes effectively.To address the challenges,we propose three methods,and the details are as follows:(1)Combining the characteristics of the brain,we propose graph modeling and frequent approximate subgraph mining methods aimed at brain medical images.These discovered patterns can assist doctors analyzing a specific disease,finding the underlying cause of the disease and latent diagnostic solutions.Additionally,the mined FASs are the basis of image classification and clustering based on FASs.In details,combining the characteristics of symmetry and connection of right and left hemispheres,we propose a graph based on the topological relations among lateral ventricles and lesion regions to represent brain medical images in the first place.Then,combining the incompatibility among different regions of brains,we develop a frequent approximate subgraph mining method based on graph edit distance to discover hidden patterns in brain medical images.At last,to improve the executive effectiveness,an approximate method is further proposed based on the greedy strategy.Experiments show that the proposed methods have higher executive effectiveness and better extension compared with comparison methods..(2)Combining the diagnostic information of brain images,we propose a brain medical image classification method based on corner detection and matching to assist doctors make diagnoses.Specifically,combining the diagnostic information that doctors pay different attention to different regions,we first develop a corner detection method based on multi-texture images,assigning corners different importance-values.Then,combining the stability and uncertainty of brain,we present an initial matched corner sequence,which exists redundancy.To end this,we employ bipartite graphs to generate a final matched corner sequence.Lastly,we devise a similarity function regarding corner matching results and classify brain medical images with the k-nearest neighbor model.Experiments show that the proposed methods have better accuracy and recall as well as higher executive effectiveness and robustness compared with comparison methods.(3)Combining existing deep CNN and recurrent neural network models,we propose a disease diagnostic method based on deep learning features from longitudinal brain medical images for more accurate computer-aided diagnosis.In details,we fine-tune an advanced deep CNN to extract features from brain medical slice images firstly,and then utilize the bag-ofwords model to integrate slice-image features,obtaining longitudinal-image features.Finally,we devise a recurrent neural network to learn the features of pathological changes from longitudinal-image features for better assist diagnosis.Experiments shown that the proposed methods can realize more accurate diagnosis...