Image-assisted Discrimination Method For Neurodevelopmental Disorders In Infants Based On Multi-feature Fusion

Premature infants have a significantly increased risk of developing severe neurodevelopmental disorders such as cerebral palsy and mental retardation due to some congenital defects at birth.Infants with abnormal neurodevelopmental outcomes behave early in behavior as abnormalities in whole body motor patterns,for example,before the full term and one month after the full term,the overall exercise sequence lacks variability and monotony,and the amplitude,speed and intensity of the exercise are small,while in the 2-5 months after the full term,small-scale moderate-speed movements in all directions throughout the body are not observed,and the movement is not smooth and uncoordinated.Based on these behavioral characteristics,doctors judge the developmental behavior of small infants and provide early intervention in small infants with abnormal assessments to improve adverse neurodevelopmental outcomes.This is very important for the treatment of neurodevelopmental disorders and to reverse their poor developmental outcomes.In this regard,we propose whether computer vision and machine learning can be used instead of human evaluation to improve the efficiency and accuracy.Firstly,this article analyzes the specific manifestations of infants’ abnormal behaviors,and tracks the infants to observe the changes in data during their movements.After analyzing the traditional Meanshift tracking algorithm,this algorithm is simple and very suitable for the single background sample video in this paper.However,considering the conditions for its judgment,a target tracking algorithm based on the attention mechanism is proposed,which is improved based on the original algorithm.To better suit the tracking of the video samples in this article.It makes up for some shortcomings of Meanshift operator and achieves good tracking results in tracking infant samples.Secondly,sample databases of multiple groups of infants were established.For feature extraction,two aspects of analysis based on motion trajectory and analysis based on human characteristics were considered.Considering the change of the motion signal,the wavelet feature information based on the motion trajectory and the power spectrum feature information based on the motion trajectory were extracted.Considering the infant’s movement,the speed feature information of the infant’s movement,the acceleration feature information of the infant’s movement,and the infant’s centroid change feature were extracted.Considering the human body characteristics,the quantitative characteristics of the infant centroid was extracted.Finally,this paper proposes a framework for detecting abnormal behaviors of infants based on multi-feature fusion.From tracking infant targets to extracting and analyzing multiple groups of features,and the results of the classification are comprehensively judged by data weighted fusion,so as to conclude whether the infant’s neurodevelopment is normal andwhether early intervention is needed.Among them,in the process of detecting infants,the traditional research on infants’ abnormal behavior often studies the infant as a whole.This article considers dividing the infant into five parts: left hand,right hand,left lower limb,right lower limb and whole body.Study these five parts separately,feature extraction and classification,and finally make a comprehensive judgment.This improves the sensitivity and specificity of various parts of the infant’s body,and also gives stronger data support for the final comprehensive judgment.