A novel two-staged deep learning based workflow for analyzable metaphase detection

Abstract

In the field of cytogenetics, chromosome image analysis plays a critical role in the diagnosis of various genetic disorders and cancers. As the gold standard, chromosome image analysis focuses on metaphase images, as all chromosomes become distinctly visible during this phase of cell mitosis. However, not all the detected metaphases are suitable for karyotyping due to chromosome stickiness. Therefore, cytogenetics selects the analyzable metaphase images by exploring all the samples. Since this task is very time-consuming, efforts have been made to automate the analyzable metaphase selection step of karyotyping. Nevertheless, these studies have drawbacks, such as utilizing hand-crafted features, requiring expert-provided candidate metaphase cell regions for classification, and demanding 100X magnified microscopic images. In this study, a novel two-staged deep-learning based workflow is presented for automatic detection of analyzable metaphase cells. The proposed model operates on 10X magnified images while omitting traditional image processing steps that are used in state-of-the-art methods for determining candidate metaphases. In the first stage, Faster R-CNN method with Inception V2 architecture is used to directly identify analyzable metaphases in specimen slide images. The second stage involves analyzable metaphase image verification based on VGG19 model. To assess the success of the proposed method, we also evaluate two state-of-the-art techniques: YOLOv8 and EfficientNet algorithms. The results obtained with different combination options achieved true positive rates up to 98.5% while reducing false positive rates to 0.0001. This study is the first to introduce an end-to-end deep learning-based approach for analyzable metaphase detection. Comparing it with existing literature, it becomes evident that the proposed approach achieves an optimal balance between true positive rates and false positive rates, maintains efficient processing speed, and eliminates the need for high-magnification image acquisition. All these factors play imperative roles in the clinical usage of the system.