Evaluation of an Artificial Intelligence-Assisted Metaphase Finder for Clinical Cytogenetic Application

Background/Objectives: Clinical cytogenetics has benefited from artificial intelligence (AI) integration into digital karyotyping systems for segmentation and classification of chromosomes. Metaphase image identification and capture utilize traditional metaphase finder classifiers whose specificity can be improved, especially where lowered detection thresholds for low-mitotic-index specimens result in higher debris selection. Use of AI to reliably detect metaphases at 10× magnification may resolve this issue. This study compares a new AI-based metaphase finder algorithm to the conventional non-AI classifier. Methods: G-banded slides from 105 normal and abnormal peripheral blood (PB) and bone marrow (BM) specimens were scanned at 10× magnification and images were captured at 100× using the HiBand system (Applied Spectral Imaging), once with the automatic metaphase finder and again using the AI-assisted algorithm. The number of real metaphases versus debris, for up to 100 captured images per slide, and quality metrics were compared using both applications. Results: True metaphase detection in BM specimens averaged 77.3 ± 21.6 versus 82.5 ± 19.9 (6.3% increase) for normal cases and 72.1 ± 18.5 versus 89.7 ± 14.2 (14.6% increase) for abnormal cases comparing the non-AI and AI classifiers, respectively. There was no difference for normal PB cases (97.6 ± 1.5 versus 97.1 ± 2.9) and a 2.6% decrease for abnormal PB cases (97.9 ± 2.2 versus 95.4 ± 4.2) between the two classifiers. All cases demonstrated comparable image quality. Conclusions: This study demonstrates that the AI-assisted metaphase finder improves true metaphase capture compared to the non-AI classifier for BM while maintaining high fidelity for PB specimens. These findings contribute to the expanding literature on integration of AI in clinical cytogenetics.