Research On MR Image Processing Technologies Via Imitating The Mechanisms Of Visual Context

To make a diagnostic decision,doctors rely on a list of perceptual and cognitive behaviors for reasoning.A doctor would first hypothesize a patient with unknown health problems,and then,verify the hypothesis through observing the patient’s magnetic resonance(MR)scans.After rounds of inference,the doctors eventually draw the conclusion only if they have eliminated information ambiguity.By contrast,although existing computer-aided diagnosis algorithms can fully mine the correlation between image features and diagnostic conclusions,these algorithms cannot deal with information ambiguity as doctors do.Therefore,they cannot be effectively embedded in the evidence-based diagnosis workflow.For unveiling how a doctor’s visual system eliminates information ambiguity,this dissertation presents a systematical and thorough investigation.Considering the complexity of diagnostic activities,this manuscript only focuses on the primary visual functions of a doctor’s individual activities,exploring how visual context eliminates information ambiguity.What follows,the author attempts to imitate such mechanisms to build a ”computational vision model” in guiding to design ”computer vision algorithms” that can adaptively cope with the ambiguity in MR images.The author’s five contributions are summarized as follows:(1)This dissertation introduces the function of visual context in three different aspects,including ”human behavior performance”,”neural signal processing mechanism” and”algorithm process”.To simulate the mechanisms of visual context,this dissertation establishes a ”context-based computational vision model” that adaptively introduces constraints and eliminates ambiguous solution utilizing contextual information.It avoids the shortage of conventional computational vision model that relies on manually specified constraints.(2)This dissertation proposes context-aware superpixel(CASP)to generate image context,which is analogous to imitate the formation of visual context.CASP adaptively generates superpixel-amount,avoiding the limitations of manual initialization of traditional algorithms.In the experimental results of segmenting brain tissues,CASP is better than traditional algorithms and is more robust against noise corruption.Besides,this dissertation proposes structural entropy,which is used to measure the cost of superpixels in coding image structure.Supported by the results,CASP achieves the smallest structural entropy,which meets theoretical expectations.Furthermore,using Pearson’s correlation analysis,this dissertation claims a strong association between intensity coherence and segmentation accuracy.(3)This dissertation proposes the multi-kernel filter,which extends bilateral filtering kernel w.r.t.filtering adaptivity.The multi-kernel filter automatically generates the parameter of range kernel,avoiding the limitations of manual initializing the range kernel of bilateral filtering.The experimental results demonstrate that the filter kernel of the multi-kernel filter is more robust than BF when it is used to smooth noise out.Especially in the task of smoothing out noise from MR images,the multi-kernel filter achieves better quantitative performance than three traditional algorithms for coping with the non-stationary noise.(4)This dissertation introduces MKF and proposes a spatially adaptive phase correction method for resolving the challenges of large variability in signal-to-noise ratio and the noise level.Compared with the four classic filters,MKF demonstrates the superiority for filtering performance demonstrated by our results.In addition,the data obtained by phase correction technology allows estimating brain microstructure,such as fractional anisotropy,more accurately.(5)This dissertation proposes RAMBundles that integrate the geometric features of fiber context and the spatial feature in a local receptive field for classifying intractable fiber bundles.From the results,RAMBundles significantlly increased the recall rate of intractable fiber bundles,and it can be observed that false-positive fibers generated are closer to the true positive ones than the traditional algorithms.