Metaphase

Feb 23, 2023

Abstract Chromosomes are a principal target of clinical cytogenetic studies. While chromosomal analysis is an integral part of prenatal care, the conventional manual identification of chromosomes in images is time-consuming and costly. This study developed a chromosome detector that uses deep learning and that achieved an accuracy of 98.88% in chromosomal identification. Specifically, we compiled and made available a large and publicly accessible database containing chromosome images and annotations for training chromosome detectors. The database contains five thousand 24 chromosome class annotations and 2,000 single chromosome annotations. This database also contains examples of chromosome variations. Our database provides a reference for researchers in this field and may help expedite the development of clinical applications. Background & Summary The human cell has one pair of sex chromosomes and 22 pairs of other chromosomes. Abnormalities in the total number or structure of chromosomes are referred to as chromosomal aberrations and are the leading cause of genetic disorders 1 . The conventional sampling method is amniocentesis, during which amniotic fluid from the uterus is aspirated under sonographic guidance. Approximately one in 150 babies have chromosomal abberations 2 . Common chromosomal aberrations occur on chromosomal pair 13 (trisomy 13), which is associated with Patau syndrome, pair 18 (trisomy 18), which is associated with Edwards syndrome, and pair 21 (trisomy 21), which is associated with Down syndrome. According to the National Center for Biotechnology Information, these chromosomal aberrations cause 50%–60% of early miscarriages. Karyotyping is clinically important in prenatal genetic diagnosis 3 . Karyotyping is a diagnostic method in which characteristic dark and light bandings of chromosomes are visualised on images for examination by physicians or senior technicians. Abnormality is determined according to the number and structure of abnormal chromosomes and sex-related chromosomes. The procedure typically takes approximately 20 min for an experienced examiner. The examiner needs to sort, cut, orient, and rearrange the mapping of a raw chromosomal cell, and at least four chromosomal images need to be processed for an individual subject to ensure a correct diagnosis. Chromosomal analysis is labour intensive and is an urgent issue because of increasing shortages of medical manpower. Automated chromosome classification systems are scarce. Most current systems are based on artificial intelligence (AI) approaches involving machine learning and deep learning 4 , 5 , 6 , 7 . Earlier studies on chromosome classification were based on segmenting overlaps and adherent chromosomes and employed conventional methods like border detection 8 , 9 , the watershed method 10 , and straightening of bent chromosomes 11 , 12 . These methods depended heavily on image preprocessing, resulting in distorted chromosome features that could result in misdiagnoses. Recent research in this field has a growing preference for chromosome prototypes over preprocessing. Chromosomes are classified by basically one of two approaches. The first approach involves the analysis of single chromosomes. This requires a human examiner, takes substantial time and effort, and is often complemented by background image segmentation and noise suppression 13 . Convolutional neural networks (CNNs) 14 , 15 may be used for classifying images; however, the accuracy is unsatisfactory due to low data volume. This approach, due to its repetitiveness and the variability of chromosome features, has limited clinical application. The second approach involves the analysis of original images by using deep learning–based object detection models 16 , 17 , 18 , 19 , 20 to identify and classify chromosomes. For example, DeepACEv2 21 requires no manual preprocessing and uses object detection as the backbone to frame and classify individual chromosomes, and this is followed by final confirmation and manual editing by a human examiner. This approach is clinically more applicable. In the literature, chromosome images are relatively easy to identify and classify from chromosome images. Despite the simplicity of these images, an examiner must spend substantial time and effort to identify the chromosomes. An automatic chromosome recognition system is essential for handling more difficult images for better clinical application. The application of AI models for medical imaging is constrained by the complexity of medical images. In a clinical setting, an incomplete AI model would not be practically useful and may even decrease staff productivity. In the event of an incomplete database, the trust of experts and patients cannot be gained 22 . Many examples of AI in medical research require a large database to improve the credibility and stability of the AI model 23 , 24 , 25 , 26 , 27 . We have developed here a detector called the ‘Automated Chromosome Detector Based on Metaphase Cell Images Using Deep Learning’ that is capable of locating and classifying chromosomes in images. The images used in this study have more chromosome overlaps and adherences than those used in other studies. Chromosome overlaps and adherences can be confusing for specialists. A probabilistic two-stage algorithm was adopted to improve chromosome detection accuracy. The method was trained and validated using data from 5,000 chromosomal images of fetuses. High accuracy (98.88%) was achieved—higher than that achieved by experienced specialists. The chromosomal images and annotations used to train the detector have been provided in this study. This is the first publicly available large database of chromosome annotations. The database contains 2,000 annotations for single chromosomes and 5,000 annotations for 24 chromosomes [Fig. 1b,c , respectively]. We also provide criteria for defining difficult images and notes from our experts on classifying chromosomes as a series of common points in the clinical recognition of difficult images. What we provide is a good benchmark dataset for researchers in this field that can expedite technical development in this application area. For example, using 5,000 annotations for 24 chromosomes, better accuracy can be achieved. These images can also help develop algorithms and expert recommendations for those images that are difficult to examine. Finally, single chromosome segmentation data can help segment chromosome overlap and adherence or to standardise the orientation of the short arms of chromosomes for examination by clinicians. Fig. 1 Example of a raw chromosome image with three annotated datasets. (a) Original chromosome image taken from fetal amniotic fluid; (b) annotation of single chromosomes; (c) annotations of 24 chromosome categories.