Research On Intelligent Assisted Analysis And Diagnosis Techn Ologies For Multi-model Medical Big Data

"Intelligence" brings human enough capacity for collecting,sorting,thinking and processing data,and the data can be absorbed to form their own knowledge structure.However,people are always so exhausted for data processing with the rapid explosion of data volume and the continuous growth of massive data sets.Therefore,they want to create multiple "agents" to help them to solve data processing problems.Artificial intelligence is a product conforming to the time.It is always expected to possess brain-like intelligence which has enough abilities of perception,cognition and decision based on the calculation and analysis of massive big data.As medical big data is closely related to human health,it has become the "bridgehead" of artificial intelligence technology development.However,due to the multimoding and high-complexity of medical big data,lots of the current methods cannot endue computers with enough "Intelligence"to deal with specific problems in medical big data analysis and diagnosis application.To overcome these problems,the entire dissertation focuses on the data analysis,feature extraction,detection and classification tasks of one-dimensional PPG signals and two/three-dimensional pulmonary nodules’ CT images,including the following parts:(1)For rapid pulmonary nodules detection task,a CV model based on improved velocity function and an improved SLIC super-pixel construction method are proposed to segment the blurring region from the suspected pulmonary nodules.The segmentation results could be more accurate to find the internal boundary and the external boundaries of the blurring region.The average depth of the segmented region is defined and computed,which could be thresholded to screen out the suspected nodules.Experimental results showed that the method could rapidly detect the suspected nodules and segment the nodules to provide pre-processing operation for the subsequent works.(2)For early-stage pulmonary nodule detection task,a novel GAN-based model,Functional-Realistic Generative Adversarial Networks model(FRGAN)is proposed to generate high-resolution nodules for the low-resolution input nodules to evaluate the malignancy.This is an exploratory work for low-resolution pulmonary nodules classification.To remain more consistent nodule representation of the input nodules,the high-resolution counterparts are generated by adding a functional loss based on attributed relational graph construction and and-or graph mining.Generalized Laplace distance is introduced to measure the similarity between the attributed relational graphs of low-resolution nodules and and-or graph template of high-resolution nodules.Based on the loss function,the generator can generate the high resolution image and predict its malignant classification after confronting the pre-trained discriminator which fed with the original high resolution image.Experimental results demonstrated that the proposed framework could improve accuracy in the task of the low-resolution pulmonary nodules detection.(3)For false positive reduction in pulmonary nodules detection,MR-Forest,a deep decision framework based on multi-ringed scanning,is proposed to relieve high computation and storage consumptions,which are always paid for the more complicated deep neural network frameworks.A series of spherical multi-scale representations are scanned with order ringed facets(ORFs),instead of the spatial sampling of the convolution feature.Moreover,the level-wise automatic growth mode of cascade forest is used in the deep decision framework for multi-scale spherical features classification.The spherical feature representations explicitly encode the Mesh-LBP based spherical texture feature,and the attraction-repulsion based fitting deformation feature and the spatial voxel position feature.The performance of proposed framework was verified on a benchmark data set and a merged dataset.Compared with those of state-of-the-art methods,it could give a competitive performance of false positive reduction with lower computational and storage resource consumption.(4)For PPG-based physiology-mental status detection task,ArvaNet,a deep recurrent neural network based on spatio-temporal graph inference,is proposed to recognize negative emotion and mental state for all-weather mental state monitoring.In this study,film clips with reference labels were used to arouse the emotions of the subjects,and their real-time emotions were inferred by analyzing the subtle waveform structure in multiple PPG fusion signal of a single cycle.Subsequently,integrating temporal factor,the mental status of the subjects was deduced by observing the difference of continuous emotional evolutions.To improve the influence of subtle waveform changes on the inferred results,a local attention mechanism based on wall shear stress was introduced.All the volunteers obtained the reference value of mental state by using DASS-21 scale and eliminated the volunteers with the tendency of extreme and severe depression and anxiety.Compared with the referenced and inferred labels,the proposed framework are demonstrated that it can provide an efficient technical support for mental state monitoring.