Fig 7 - uploaded by Xu Cao
Content may be subject to copyright.
Two examples of failed cases. (a): blue boxes denote connecting crossing-overlap; (b): red circles denote connecting non-overlapping parts.
Similar publications
Citations
... The proposed denoising and segmentation models were used to accomplish chromosome separation. Previous studies used object detection methods 12,13 and semantic segmentation 9,14,15,19,51 for chromosome separation. DNN-based chromosome detection methods aim to enumerate chromosomes rather than produce karyograms. ...
... 14 With a reconstruction strategy, a semantic segmentation model was applied to recognize overlapping and nonoverlapping chromosomal parts and reconstruct intact chromosomes. 19 However, in the practical application of clinical chromosome segmentation, chromosomes are located randomly within a metaphase on the slide, requiring a segmentation model capable of solving complicated situations for each metaphase. Of note, semantic segmentation models cannot simultaneously solve all the aforementioned problems. ...
Context.—
Conventional karyotype analysis, which provides comprehensive cytogenetic information, plays a significant role in the diagnosis and risk stratification of hematologic neoplasms. The main limitations of this approach include long turnaround time and laboriousness. Therefore, we developed an integral R-banded karyotype analysis system for bone marrow metaphases, based on deep learning.
Objective.—
To evaluate the performance of the internal models and the entire karyotype analysis system for R-banded bone marrow metaphase.
Design.—
A total of 4442 sets of R-banded normal bone marrow metaphases and karyograms were collected. Accordingly, 4 deep learning–based models for different analytic stages of karyotyping, including denoising, segmentation, classification, and polarity recognition, were developed and integrated as an R-banded bone marrow karyotype analysis system. Five-fold cross validation was performed on each model. The whole system was implemented by 2 strategies of automatic and semiautomatic workflows. A test set of 885 metaphases was used to assess the entire system.
Results.—
The denoising model achieved an intersection-over-union (IoU) of 99.20% and a Dice similarity coefficient (DSC) of 99.58% for metaphase acquisition. The segmentation model achieved an IoU of 91.95% and a DSC of 95.79% for chromosome segmentation. The accuracies of the segmentation, classification, and polarity recognition models were 96.77%, 98.77%, and 99.93%, respectively. The whole system achieved an accuracy of 93.33% with the automatic strategy and an accuracy of 99.06% with the semiautomatic strategy.
Conclusions.—
The performance of both the internal models and the entire system is desirable. This deep learning–based karyotype analysis system has potential in clinical application.
... In 2020, Cao et al. [6] proposed an automatic segmentation method that first segments and then splices for overlapping and adherent chromosomes, with an overall accuracy of 82.60%. The segmentation results are shown in Fig. 1. ...
... The segmentation results are shown in Fig. 1. Based on the segmentation results of the method [6], it can be concluded that this method can segment overlapping chromosome images in most cases. However, in some special overlapping cases, when two or more overlapping regions in the chromosome image are too close to each other, the non-overlapping regions that do not belong to the overlapping region are matched, and the chromatids in the chromosome image cannot be completely separated. ...
... The central coordinates of the non-overlapping regions within the combination were determined in Step2. In geometry, two points can determine a line segment, so four central coordinates can determine C 2 4 ¼ 6 line segments, and each line segment can be expressed as l j [(x j1 , y j1 ), (x j2 , y j2 )], j 2 [1,6]; ...
Chromosome images are commonly used in karyotype analysis to diagnose chromosomal diseases. However, there are often chromosome adhesion and overlaps in chromosome images, so effective chromosome segmentation is conducive to smooth karyotype analysis. To date, some progress has been made in automatic chromosome segmentation, and existing methods can be used to segment overlapping chromosomes in most cases. However, when two or more overlapping regions are too close to each other in the image of overlapping chromosomes, the existing segmentation methods adjust the non-overlapping regions that do not belong to the overlapping region, resulting in incomplete segmentation of chromatids. Therefore, we use a heuristic algorithm to solve this problem from the point of view of mathematics and geometry to improve the segmentation of overlapping chromosomes. Starting from chromosome images, the existing problems and solutions are explained and displayed in the way of visualized interpretable image features, which helps to better understand the algorithm. Our method achieves 92.86% splicing accuracy and 90.44% overall segmentation accuracy on open datasets. The experimental results show that our method can effectively improve the problem of incorrect chromosome segmentation when two or more overlapping parts of overlapping chromosomes are too close to each other. It can accelerate the development of artificial intelligence in computational pathology and provide patients with more accurate medical services.
... In the second stage they separate the chromosomes instances using an image processing algorithm. On top of segmenting where the individual chromosomes are, Cao et al. (2020) also reconstructs an image of the separated chromosomes. ...
A difficult step in the process of karyotyping is segmenting chromosomes that touch or overlap. In an attempt to automate the process, previous studies turned to Deep Learning methods, with some formulating the task as a semantic segmentation problem. These models treat separate chromosome instances as semantic classes, which we show to be problematic, since it is uncertain which chromosome should be classed as #1 and #2. Assigning class labels based on comparison rules, such as the shorter/longer chromosome alleviates, but does not fully resolve the issue. Instead, we separate the chromosome instances in a second stage, predicting the orientation of the chromosomes by the model and use it as one of the key distinguishing factors of the chromosomes. We demonstrate this method to be effective. Furthermore, we introduce a novel Double-Angle representation that a neural network can use to predict the orientation. The representation maps any direction and its reverse to the same point. Lastly, we present a new expanded synthetic dataset, which is based on Pommier's dataset, but addresses its issues with insufficient separation between its training and testing sets.
A difficult step in the process of karyotyping is segmenting chromosomes that touch or overlap. In an attempt to automate the process, previous studies turned to Deep Learning methods, with some formulating the task as a semantic segmentation problem. These models treat separate chromosome instances as semantic classes, which we show to be problematic, since it is uncertain which chromosome should be classed as #1 and #2. Assigning class labels based on comparison rules, such as the shorter/longer chromosome alleviates, but does not fully resolve the issue. Instead, we separate the chromosome instances in a second stage, predicting the orientation of the chromosomes by the model and use it as one of the key distinguishing factors of the chromosomes. We demonstrate this method to be effective. Furthermore, we introduce a novel Double-Angle representation that a neural network can use to predict the orientation. The representation maps any direction and its reverse to the same point. Lastly, we present a new expanded synthetic dataset, which is based on Pommier’s dataset, but addresses its issues with insufficient separation between its training and testing sets.KeywordsKaryotypingDeep learningSemantic segmentationInstance segmentationRepresentationInvariance
