Introduction:
Determining the karyotype of a patient is crucial to understanding their genetic makeup and identifying disease-related features of the patient’s chromosomes. This is done by analyzing cells in the metaphase step of the cell cycle, because this is when chromosomes are compact and can clearly be distinguished. In the clinical lab, microscopy slides are prepared to maximize the prevalence of cells in metaphase but identifying where these metaphase cells are on the slide is a time-intensive task. This study aims to understand if the metaphase finding process can be improved by using deep neural networks (DNNs), so analysis and patient karyotype results can be acquired more efficiently and with fewer failed slides. This could ultimately allow for patient needs to be addressed in a timelier manner.

Methods:
In this study, slides were scanned using the MetaSystems Metafer slide-scanning system. This specific application requires an LED light source for brightfield microscopy, a Zeiss Z2 motorized microscope, a 10x objective, and a 12-megapixel camera for hardware. The software required consists of the appropriate licenses and installed applications from MetaSystems. 40 G-banded bone marrow slides prepared for karyotyping were scanned using both the Metaphase DNN classifier and the MSearch classifier. The Metaphase DNN classifier applies an object detection DNN on full fields of view to detect possible metaphases first. Then, it applies 2 classification DNNs on these possible metaphases to give a quality rank based on 1) metaphase spreading and 2) chromosome length. The MSearch classifier is the most current and widely validated metaphase finding classifier in the cytogenetics field in the United States which uses feature-based classification, and it was run with 5 different sensitivities: 6, 7, 8, 9, and 10. Metaphases found with either method were scored as either analyzable or non-analyzable, so that each tested condition had true positive metaphases, false positive metaphases, and the respective quality scores for each metaphase.

Results:
With this study, we show that introducing DNNs significantly improves the automatic identification of high-quality metaphase cells in bone marrow samples when compared to the widely used, clinical standard MSearch in Metafer. We found that when utilizing DNNs, the metaphases found by the DNN are of higher quality and the false positive rate is lower than when using the most specific setting in the current standard. DNNs also increase the number of analyzable metaphases found when compared to the most sensitive settings used in the current standard, leading to fewer missed metaphases.

Conclusion:
These results demonstrate the advantages of using DNNs when detecting metaphase cells on a microscopy slide compared to the current standard, identifying more metaphase cells and with greater confidence that they are analyzable metaphase cells. Implementing DNNs for detection of metaphase cells has the potential to further reduce technologists’ time spent on the initial task of finding and capturing metaphase cells and can allow for an increase in their time spent on more challenging cytogenetic tasks during analysis. This can both decrease the time it takes to deliver results to a patient and enhance job satisfaction for technologists by allowing them to engage in more stimulating work. Additionally, using DNNs can potentially reduce the number of samples where no metaphases were found, reducing the amount of wasted lab resources and increasing patient care capacity. While these results were found using bone marrow samples, expanding this strategy to peripheral blood samples may also prove fruitful especially for constitutional cases where analyzing high-quality metaphase cells is sufficient for a clinical result.