Research On The Application Of Deep Learning In Human Karyotype Analysis

Karyotype analysis is an important tool for diagnosing the presence of chromosomal disorders in humans.Traditional karyotype analysis relies on manual processing by domain experts,which is time-consuming and laborious to operate.To solve such problems,deep learning-related techniques are applied to karyotype analysis,and a series of deep learning models are designed to segment and categorize instances of chromosome microscopic images.Finally,based on the theoretical study,a karyotype analysis system is designed and implemented,which can complete automatic karyotype analysis of chromosomes to assist the manual operation of domain experts and greatly improve the work efficiency.To solve the problem of karyotype analysis,the following research work is carried out in this paper:(1)Firstly,to address the problems of impurities and high similarity between chromosomes and background in chromosome microscopic images,U-net,a semantic segmentation algorithm,is used to extract pixels belonging to chromosomes in chromosome images,which in turn results in background removal and de-cluttering.Due to the small number of privately acquired chromosome datasets,data enhancement algorithms such as rotation,translation,and color dithering are proposed to expand the datasets.Based on this,the Hybrid Task Cascade architecture(HTC)is used to identify and segment chromosome instances,and the feature pyramid structure in the model is improved by proposing the use of path augmented feature pyramid structure(PAFPN)instead of feature pyramid structure(FPN)to retain the shallow feature information and further improve the accuracy of localization and segmentation.(2)In order to solve the problems that the training speed and the number of parameters of HTC model are large,and the effect of segmenting overlapping regions in chromosome images is not good,the classical instance segmentation model Mask R-CNN is proposed for instance segmentation of chromosome images.Based on this model,a chromosome image synthesis algorithm is designed to augment the dataset;a new IoU metric is introduced to improve the quality of positive samples for target detection;the Soft NMS method is introduced for overlapping chromosome clusters to improve the screening of candidate frames and the suppression of overlapping frames;the CBAM attention mechanism and MaskIoU are introduced in the semantic segmentation branch of Mask R-CNN to improve the semantic segmentation of the model module in the semantic segmentation branch of Mask R-CNN to improve the pixel loss problem in the semantic segmentation of the model.The experimental results show that the improved Mask R-CNN model is able to segment chromosome instances in overlapping regions of the image better while ensuring the accuracy of segmentation,and the final segmentation mAP achieves an accuracy of 78.52%.(3)The VAN model utilizes the large kernel attention mechanism to achieve adaptive and remotely associated self-attentiveness with reduced computational cost,while taking into account spatial and channel adaptation.Thus it can better focus on the chromosome feature information in the classification dataset and improve the higher classification accuracy.Therefore,VAN is chosen as the underlying classification model architecture in this paper.Since the texture information and chromosome size of chromosomes are important bases for chromosome classification,a texture enhancement algorithm is performed for classified images with blurred texture in the classification dataset to expand the contrast between light and dark in the images.Then the chromosome classification dataset is expanded using random horizontal and vertical flipping,random panning and random clipping operations,especially for the relatively small number of X and Y datasets.Finally,a predictive label correction algorithm based on prior knowledge is proposed to correct the classification results,and finally the average accuracy of chromosome 24 classification reaches 98.26%.(4)Finally,this paper uses the Visual Studio development platform based on the actual requirements for the construction of the karyotype analysis system,combined with MySQL database and the algorithms of deep learning involved above to realize the software development.After analyzing the system requirements,a complete karyotype analysis system for chromosomes is designed and implemented.The main functions of the system include image input,background purity,texture enhancement,automatic segmentation,auxiliary segmentation,automatic classification,auxiliary classification and karyotype result generation.The use of the system enables automated karyotype analysis and visualization of the result graphs at each stage of analysis,aiding physicians in better and faster analysis and diagnosis.