Research On Functional Magnetic Resonance Image Retrieval For Brain Disease Analysis

The prevalence of brain diseases is increasing and seriously affects human life and health.Therefore,the intelligent diagnosis of brain diseases is of significant research value.With the widespread use of functional Magnetic Resonance Imaging(f MRI)in brain diseases,intelligent assisted diagnosis of brain diseases based on f MRI has become possible.Medical image retrieval technology,which enables the retrieval of images and diagnostic cases with the most similar pathological features to the current image from an existing case database,thus assisting doctors in diagnosis and treatment,is one of the important applications of computer-aided diagnosis.However,most of the current retrieval efforts for f MRI of brain diseases are still focused on conventional medical image retrieval methods,and there are two major problems.On the one hand,the advent of the big data era has led to an explosive growth in f MRI data,and when extracting high-dimensional features for retrieval,the time complexity and spatial complexity of the retrieval method is very high.On the other hand,in reality,large medical image databases are often built from multiple data centers,and existing approaches often simply combine multi-center data into one dataset,ignoring the heterogeneity of different data sources(i.e.differences in data distribution),which to some extent affects the generalization performance of the model.To address the above issues,a series of deep learning f MRI data retrieval methods are proposed in this thesis and experimentally validated on a large f MRI dataset.The main research elements are summarized as follows:(1)A deep hash-based f MRI retrieval method is proposed for fast and accurate retrieval of functional magnetic resonance images.First,points in Hamming space that are sufficiently distant from each other are called hash centers,and then global similarity learning is performed on the hash representations of the samples so that the hash representations of similar samples are close to the same hash center.To further improve the quality of the discrete hash codes while making better use of the semantic information of the data,a hash coding layer and a classifier are set up in the network,the former is used to ensure strict discrete constraints in forward propagation,the classifier classifies the discrete hash codes,and the gradient in backward propagation is passed to the previous layer intact,by optimizing the classification loss and thus directly learning the binary hash codes also avoiding the optimization NP hard.(2)A multi-site f MRI hash retrieval method based on single-source-domain adaptation is proposed.A multi-way parallel retrieval framework containing multiple sub-models is designed for the multi-site distribution discrepancy problem,and in each sub-model the data distribution is aligned according to optimal transport theory minimizing the transport cost between the source and target domain distributions.Using an alternating optimization approach,the training process of the hash function in the first work is embedded in the learning of the optimal transfer matrix in each sub-model,which reduces the impact of generalization due to differences in the multi-site data distribution while efficiently learning the hash codes.(3)A multi-site f MRI hash retrieval method based on graph convolution and stable learning is proposed.First,the graph convolutional neural network can update and aggregate the feature representations of each brain region,which in turn captures the topological features of the brain network,making it possible to fully extract information from the brain network data.To further address the issue of multi-site data distribution discrepancies,according to the theory of stable learning,sample weighting is used to eliminate spurious associations and make the model focus more on the essential features that are only causally related to the category,so that even with multi-site data,the model only extracts the most essential features of the sample,thus improving the model generalization performance.The sample weighting is achieved by weighting the classification loss.The hash learning aspect is still performed by setting hash centers in Hamming space to learn the global similarity of the hash representation of the samples,while the classification loss introduced can further mine the semantic information of the data and accelerate the model convergence.The combination of the above two parts improves the performance of cross-site retrieval.