Research On Chromosome Karyotyping Methods Based On Deep Learning

Chromosome karyotyping is a widely used method for the diagnosis of genetic diseases,which involves the separation and classification of chromosomes in the images of cells during metaphase of cell division.The non-rigid nature of chromosomes causes them to have different poses at different times and in different cells,and the problem of overlapping chromosomes adds significant difficulty to karyotyping.Currently,karyotyping still heavily relies on manual or semi-automated methods,which poses a great burden on the staff.Therefore,this article focuses on the research of overlapping chromosome segmentation and chromosome classification on metaphase images.The main contributions are as follows.:(1)Aiming at the problem that the size of overlapping regions of chromosomes is different,and the segmentation model cannot effectively extract context information,this thesis constructs a two-stage overlapping chromosome segmentation model SC-Net,which first uses the semantic segmentation network SC-UNet++ to divide overlapping chromosomes into overlapping regions and non-overlapping regions.The network obtains targeted multiscale contexts by fusing global and local contexts,and captures category context information through context prior to enhance the model’s understanding of the overlapping relationship of chromosomes and improve segmentation continuity.Then,these regions are spliced into a single chromosome using a chromosome instance reconstruction algorithm.Experiments were conducted on the publicly available ChromSeg dataset.The results showed that the average crossover ratio of SC-Net for segmenting overlapping regions of chromosomes was 83.4%,with an overall accuracy of 92.8%,which is 2.4%and 2.3%higher than the existing best methods.(2)Aiming at the problem of insufficient samples in metaphase images and high similarity of chromosome types,this thesis proposes a simulated data expansion algorithm to expand the training set and constructs a chromosome classification network Mask ACK based on an instance segmentation framework.Firstly,the feature learning of the network is embedded into a convolutional binary tree to achieve fine-grained classification.Secondly,a loss of classification consistency is introduced to promote mutual learning among multiple classifiers to reduce prediction differences.Finally,based on the fact that chromosomes often appear in pairs,the type re-assignment algorithm is proposed to reclassify network predictions.Experiment on private datasets.The results show that the accuracy of the Mask ACK classification reaches 94.9%,which is 7.9%higher than the SOTA method.(3)Aiming at the problem that Mask ACK is not effective in identifying overlapping chromosomes in metaphase images,this thesis proposes a cascade model ACK-Net that combines overlapping chromosome segmentation.It first uses a lightweight classification network to extract and remove overlapping chromosomes in metaphase images.Then it applies SC-Net to separate overlapping chromosomes and generate new metaphase images.Finally,it applies Mask ACK for karyotyping.The experimental results show that the classification accuracy of overlapping chromosomes improves by 5.2%compared to Mask ACK.The IoU of overlapping chromosomes improves by 19.05%,demonstrating the effectiveness of ACK-Net.