Automatic detection of sister chromatid exchanges using machine learning models and image analysis algorithms

cell culture

Wild-type (WT) TK6 cells were provided by Dr. Koji Hirota (Department of Chemistry, Graduate School of Science, Tokyo Metropolitan University). BLM^−/− TK6 cells were provided by Dr. Masamitsu Honma (National Institute of Health Sciences). These cells were cultured at 37 °C in RPMI medium (Wako) supplemented with 5% horse serum (Gibco), penicillin/streptomycin mixture (Nacalai Tesque), 2 mM l -glutamine (Nacalai Tesque), and 1.8 mM sodium pyruvate.

SCE plate preparation and image acquisition

Metaphase chromosome spreads were prepared from WT TK6 cells; BLM^−/− TK6 cells. Cells were cultured in medium containing 10 μM BrdU for two cycles and treated with colcemid for the final 135 min. Cells were harvested and treated with 75 mM KCl for 15 min, followed by fixation with methanol-acetic acid (3:1) for 30 min. The cell suspension was dropped onto a wet glass slide and allowed to air dry. Cells on the slides were incubated for 20 min with 10 μg/mL Hoechst 33,258 (Sigma) in phosphate buffer (pH 6.8) and incubated with McIlvaine solution (164 mM Na₂HPO₄16 mM citric acid. pH7.0). Cells were exposed to black light (352 nm) at a distance of 1 cm for 25 min, incubated in 2× SSC (300 mM NaCl, 30 mM sodium citrate) at 58 °C for 20 min, and then stained with 5% Giemsa (Wako) for 10 min. Chromosome images were collected using a DS-Fi3 camera (Nikon) mounted on an inverted microscope (ECLIPSE Ni; Nikon) with a 100 × NA 1.49 objective.

Extract the chromosome image from the acquired image

To create the dataset for MASK R-CNN, each chromosome across the entire image was annotated pixel by pixel. Mask R-CNN. Excellent performance in chromosome karyotyping tasks.¹⁷outputs the class name, bounding box, and segmentation layer. The bounding box of the output layer is a rectangle along the two axes of image coordinates. To prevent interference from neighboring chromosomes, the bounding box was adjusted to maximize the chromosome area within the bounding box. In the “chromosome mask” output of the output layer, we used OpenCV’s minAreaRect function to adjust the chromosome detection window (Supplementary Figure 1). cocoa annotator¹⁸A web-based image annotation tool was used to construct the dataset. Both single and multiple chromosome classes were defined to exclude duplicate chromosomes from the analysis. Applied transfer learning using ResNet50¹⁹ Backbone with Feature Pyramid Network (FPN)²⁰ Pre-trained on ImageNet. Evaluation metrics included COCO AP (average accuracy) metrics, namely AP50, AP75, and AP for single chromosomes. Since the main objective was to extract true single chromosomes, an additional metric was introduced: AR (Analyzable Ratio). AR is calculated as follows:

where \(\:{N}_{posi}\) is the number of true positive chromosomes, \(\:{N}_{All}\) is the total number of chromosomes detected by the Mask R-CNN model. In this study, the detection threshold is set to a confidence score of 0.6.

vision transformer

ViT is an extension model of Transformer^{twenty one}widely used in natural language processing tasks.^22,23into the field of computer vision. ViT has also been successfully applied in many fields, such as in the financial field for asset price and direction prediction.^24,25and earth and environmental sciences for tasks such as land cover classification.^26,27. Transfer learning was performed using the ViT-Base model pre-trained on ImageNet-21k (14 million images, 21,843 classes). Chromosome images were resized to 224 × 224 pixels and normalized to match the RGB mean and standard deviation of ImageNet-21k. Training was performed over 10 epochs with a learning rate of 0.0003, and the weights decayed by 0.7 per step. To improve accuracy, the dataset was randomly shuffled and the model was retrained 15 times.

ViT accuracy was calculated using one of the following two formulas:

or

where \(\:{N}_{ViT}\) number of correct data points according to the ViT model; \(\:{N}_{man}\) is the manually selected number of chromosomes, \(\:{N}_{All}\) is the total number of chromosomes detected by the Mask R-CNN model. Equation (2) was used in Figure 3 to obtain the simple accuracy of the ViT model. In Table 2, we used (3) to obtain the accuracy of the ViT model combined with the Mask R-CNN model.

SCE counting algorithm

A schematic diagram of the algorithm is shown in Figure 4.

1. (I)

   Chromosome detection

This process used grayscale chromosome images that were masked based on the mask layer obtained from Mask R-CNN. One of the chromatids was identified by extracting the three pixels with the lowest pixel value in each row from the image.

1. (ii)

   Clustering of chromatid parts

To cluster the pixels extracted in (i), we applied a clustering algorithm, Density-Based Spatial Clustering for Noisy Applications (DBSCAN).²⁸. DBSCAN requires two parameters. Epsilon is a parameter that defines the maximum distance between two points for them to be considered neighbors, and the other is the minimum number of points to form a dense region (minPts). DBSCAN starts from an arbitrary starting point and searches the epsilon neighborhood of that point to form clusters containing points greater than or equal to minPts. Otherwise, the point is labeled as noise. DBSCAN is used with an arbitrary distance function and the process continues until all points are labeled. Next, we determined the reference points for each cluster obtained by DBSCAN. Specifically, the average position of all pixels within each cluster was adopted as the reference point.

1. (iii)

   center line detection

To obtain a reference line passing through the center of the chromosome, we applied a skeleton pruning based on the Discrete Curve Expansion (DCE) method.²⁹ Mask Apply to the binarized mask layer obtained from R-CNN. Skeleton pruning is a technique that converts a binary image into 1-pixel wide lines. The central reference line was determined based on a 5% cropped average of the x-coordinates of all pixels along the line generated by skeleton pruning.

1. (iv)

   Measuring replacement frequency

The number of SCEs was estimated based on the positional relationship between the reference point of each cluster obtained in (ii) and the central reference line obtained in (iii).

The accuracy of the algorithm (without ViT) was calculated as follows.

where \(\:{N}_{Alg}\) is the number of correct data points by the algorithm, \(\:{N}_{All}\) is the total number of chromosomes detected by the Mask R-CNN model.

The accuracy of the algorithm (using ViT) was calculated as follows.

where \(\:{N}_{Alg}\) is the number of correct data points by the algorithm, \(\:{N}_{sce}\) Number of SCE-positive chromosomes according to the ViT model.

A method that combines Mask R-CNN, ViT, and SCE counting algorithms

As shown in Figure 2, the input whole chromosome image was first processed using the masked R-CNN model. The excised single chromosomes were then classified by the ViT model, followed by analysis of SCE-positive chromosomes using the SCE counting algorithm. The accuracy of the combined method was calculated as follows.

where \(\:{N}_{All}\) is the total number of chromosomes detected by the Mask R-CNN model. \(\:{N}_{sce}\) Number of SCE-positive chromosomes according to ViT model, \(\:{Acc}_{ViT2}\) ViT accuracy and \(\:{Acc}_{Alg2}\) is the algorithm accuracy (with ViT).