Introduction:
Cytogenetics laboratories are experiencing a reduction in the workforce and challenges replacing highly experienced technologists. To address these challenges, there is a growing demand for computer-aided karyotyping software designed to enhance the efficiency of karyogram preparation, thereby optimizing resource allocation for higher-value tasks. Recent advancements in artificial intelligence (AI) have been integrated into these systems, significantly reducing the turnaround time required to achieve results compared to conventional digital imaging systems that rely on traditional image processing technologies. Following a small pilot evaluation reporting 92% correct segmentation and classification of cytogenetically normal peripheral blood and bone marrow chromosomes with the AI technology, a recent multicenter study showed a 46% reduction in time technologists spent karyotyping cytogenetically normal bone marrow samples with AI-based software compared to existing non-AI processes. This work expands on the previous study by comparing the two methods for karyotyping peripheral blood metaphases.

 

Methods:
G-banded slides of normal peripheral blood samples from ten patients (5 females and 5 males) were prepared by standard methods and scanned using the HiBand system (Applied Spectral Imaging). Metaphases were automatically identified at 10X magnification and images were captured at 100X. Twenty cells were selected for each case based on established laboratory analysis protocols. Cases were analyzed twice by cytogenetics technologists, once with the current non-AI software and a second time with the AI-based technology. Time required to correct and analyze each cell was recorded with both methods and compared. Statistical significance was assessed using the Wilcoxon Signed-Rank two-tailed test. A p-value lower than 0.05 was considered significant.

Results:
Two hundred cytogenetically normal peripheral blood metaphases were included in this study and a total of 9,187 chromosomes (0.1% random losses) were analyzed. Correct segmentation was achieved for 93% of the analyzed chromosomes. The average number of segmentation errors per metaphase was 3.4±2.8 (range 0-14, median 3.0). The average number of placement errors per metaphase following segmentation correction was 0.6, with a range of 0-11 errors (median 0.0). The average number of manual adjustments required by the technologist per karyogram was 3.6±2.8, with a range of 0-14 adjustments and a median of 3.0 adjustments per metaphase. No placement error was reported in 70% of the metaphases (140 out of 200). 17.5% of the metaphases (35 out of 200) required one single placement correction. The average analysis time per metaphase for both segmentation and placement was 1.6 minutes, with a standard deviation of 1.2 minutes. The average time required to analyze each case was reduced from 69±20 minutes with the non-AI process to 32±12 minutes with the AI software. This resulted in a statistically significant (p=0.005) 53% decrease in analysis time.

Conclusion:
The accuracy and performance observed for peripheral blood specimens was comparable to those reported for bone marrow cases. For both sample types, the AI technology reduces by approximately half the time necessary to review and correct karyograms, requiring an average of 4 corrections per metaphase. Accuracy in chromosome segmentation and classification is projected to further improve while algorithm models are exposed to data from additional sources. Moreover, as AI technology progresses, correct contouring and placement of abnormal chromosomes, including those with rearrangements, will further increase efficiency in the karyotyping process.