DeepKaryo-Check: a two-stage automated screening framework for chromosomal numerical and structural abnormalities in clinical karyotype analysis.

Introduction: Chromosomal abnormalities, including numerical abnormalities such as aneuploidy and structural abnormalities such as microdeletions and complex rearrangements, remain challenging to detect in a stable and reliable manner using automated methods in clinical karyotype analysis. This difficulty is primarily attributable to substantial non-rigid chromosome deformation, variability in sample preparation and imaging conditions, and the persistent scarcity of abnormal cases in real-world clinical datasets, all of which hinder the construction of generalizable morphological representations. In clinical practice, limited sample size is a common condition rather than an exception, which further intensifies the trade-off between sensitivity and specificity in automated screening systems. Therefore, achieving robust high-specificity screening under small-sample conditions is a central problem in the automated analysis of chromosomal abnormalities. In this study, we focus on cytogenetic image analysis and propose a structured modeling framework designed for clinical screening scenarios. Recent advances in multi-omics and machine learning have enabled more comprehensive characterization of disease states. Although the present work primarily focuses on cytogenetic imaging, the proposed framework provides an extensible basis for future integration with heterogeneous biological data. Methods: We propose a multi-stage automated chromosomal analysis framework, termed Results: In the independent Stage I test set of 520 karyotype images, including 260 normal and 260 numerically abnormal images, DeepKaryo-Check achieved an accuracy of 95.19%. The framework correctly classified 249 of 260 normal images and 246 of 260 abnormal images. The independent Stage II test cohort included 45 patients, comprising 29 patients without structural abnormalities and 16 patients with structural abnormalities. At the locked high-specificity threshold, the patient-level ROC-AUC was 0.9978 and the PR-AUC was 0.9963. All 29 patients without structural abnormalities were correctly classified, while 13 of 16 patients with structural abnormalities were identified, corresponding to a sensitivity of 0.8125 and a specificity of 1.000. Three patients with structural abnormalities were missed, and no patient without a structural abnormality was classified as positive. Conclusion: These findings indicate that explicitly incorporating domain-relevant geometric priors into automated analysis workflows can improve the reliability of chromosomal structural abnormality screening under sample-limited conditions.