Deep Learning-based Automatic Karyotyping for the Diagnosis of Fetal Trisomy

Wang, Chengyu (2023) Deep Learning-based Automatic Karyotyping for the Diagnosis of Fetal Trisomy. PhD thesis, University of Liverpool.

Text thesis_201406122.pdf - Author Accepted Manuscript Access to this file is embargoed until 1 August 2024. Download (10MB)

Abstract

Birth defects affect 35% of all newborns worldwide, resulting in deaths within 28 days of birth for an average of 240,000 babies annually. Trisomy is a severe congenital disorder caused by extra chromosomes in cells that can lead to a shortened life expectancy, memory impairment, delayed or impaired physical development and a distinctive facial appearance. Prenatal trisomy screening is considered essential, and the gold standard for its diagnosis is a process known as karyotyping, which identifies all chromosomes from metaphase images of a eukaryotic cell and analyses them for structural or numerical abnormalities. Karyotyping requires specialist knowledge and is time-consuming and labour-intensive. For this reason, computer-aided technology has been used for decades to investigate automated karyotyping, saving valuable diagnostic time and improving the efficiency and accuracy of diagnosis. The steps for fully automated karyotyping often include: (1) segmenting approximately 46 chromosomes from microscopic images; (2) classifying the segmented chromosomes into one of 24 categories; and (3) diagnosing abnormalities based on the structure or number of chromosomes. Since the 2010s, researchers have used machine learning methods, particularly deep learning, for breakthroughs in chromosome segmentation and classification tasks. However, most of these methods focus on single segmentation or classification tasks, and the validation datasets only include one staining technique. These studies have limitations from the clinical application perspective: (1) Fully automatic karyotyping requires combining segmentation and classification tasks to generate karyotypes from metaphase images. However, studies of separate segmentation or classification tasks do not demonstrate their efficiency in joint missions. (2) Several different staining techniques are encountered in clinical applications, and the resulting chromosomes differ significantly in banding characteristics, which is a challenge even for professionals to distinguish chromosomes. Therefore, computer-aided karyotyping tasks are expected to be validated using data from different staining techniques to validate their clinical applicability. To address the above limitations and challenges, we propose the following in this research: (1) A classification model based on Siamese structures is designed to exploit the fact that normal autosomes always occur in pairs with higher classification accuracy. (2) To meet the practical requirements of clinical application, a Fully Automatic Karyotyping Architecture (FAKA) is proposed to generate karyotype images from metaphase images. With the strategy of mapping a 3D mesh to a 2D image with a regular grid of pixels, a segmentation model improves the fineness of the boundaries of chromosome instances. (3) Based on FAKA, a Down Syndrome Detector (DSD) is proposed to detect Down Syndrome from metaphase images. The backbone network of the segmentation and classification module of DSD adopts the Transformer technology to improve the feature extraction ability. (4) The validation datasets of the proposed models are constructed from three mainstream chromosome staining techniques, making this study more consistent with clinical application scenarios. 出生缺陷影响全世界35%的新生儿，每年平均有24万名婴儿在出生后28天内死亡。三体综合征是一种严重的先天性疾病，由细胞中多余的染色体引起，可导致预期寿命缩短、记忆障碍、身体发育迟缓或受损以及面部外观与众不同。产前三体综合征筛查被认为是必不可少的，其诊断的黄金标准是一个被称为核型的过程，它从真核细胞的分裂期图像中识别所有染色体，并分析它们的结构或数字异常。染色体分型需要专业知识，并且是耗时和劳动密集型的。由于这个原因，几十年来，计算机辅助技术已被用于研究自动核型，节省了宝贵的诊断时间，提高了诊断的效率和准确性。全自动化核型的步骤通常包括：(1）从显微镜图像中分割出大约46条染色体；（2）将分割的染色体归入24个类别中的一个；以及（3）根据染色体的结构或数量诊断出异常。 自2010年代以来，研究人员利用机器学习方法，特别是深度学习，在染色体分割和分类任务方面取得了突破。然而，这些方法大多专注于单一的分割或分类任务，而且验证数据集只包括一种染色技术。从临床应用的角度来看，这些研究有以下局限性：（1）全自动的核型需要结合分割和分类任务，从蜕变期图像中生成核型。然而，对单独的分割或分类任务的研究并没有证明它们在联合任务中的效率。(2）在临床应用中会遇到几种不同的染色技术，所产生的染色体在带状特征上有很大的不同，即使对专业人员来说，区分染色体也是一个挑战。因此，计算机辅助核型任务有望利用不同染色技术的数据来验证其临床适用性。 为了解决上述限制和挑战，我们在本研究中提出了以下建议：（1）设计了一个基于连体结构的分类模型，利用正常常染色体总是成对出现的事实，具有较高的分类精度。(2) 为了满足临床应用的实际要求，提出了一个全自动核型结构（FAKA），以从蜕膜期图像中生成核型图像。通过将三维网格映射到具有规则像素网格的二维图像的策略，分割模型提高了染色体实例边界的精细度。(3) 基于FAKA，提出了一个唐氏综合症检测器（DSD），用于从偏相图像中检测唐氏综合症。DSD的分割和分类模块的骨干网络采用了Transformer技术来提高特征提取能力。(4）提出的模型的验证数据集由三种主流染色体染色技术构建，使本研究与临床应用场景更加一致。

Item Type:	Thesis (PhD)
Uncontrolled Keywords:	Karyotyping, Trisomy, Down Syndrome, Chromosome Detection, Chromosome Segmentation, Chromosome Classification, Transformer, Deep Neural Network, Convolutional Neural Network, Siamese
Divisions:	Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science
Depositing User:	Symplectic Admin
Date Deposited:	29 Aug 2023 15:23
Last Modified:	29 Aug 2023 15:24
DOI:	10.17638/03169837
Supervisors:	Ma, Fei Yu, Limin Selis, Valerio
URI:	https://livrepository.liverpool.ac.uk/id/eprint/3169837