Tokensome: Towards a Genetic Vision-Language GPT for Explainable and Cognitive Karyotyping

Artificial intelligence (AI) has achieved significant progress due to recent rapid advances in deep learning [1] techniques. Through deep learning's powerful automated feature extraction capacities--enabling detection of nuanced multidimensional patterns in medical images beyond human discernment--AI holds considerable potential to unlock substantial improvements in medical imaging domain [2][3]. However, integration of AI technologies into real-world clinical settings remains severely limited. A major obstacle is the predominant opacity of state-of-the-art AI systems, which frequently manifest as inscrutable "black-box" models that provide no actionable evidence or confidence metrics to justify their output decisions and predictions [4]. This lack of model explainability or interpretability not only hinders scientific understanding of system behavior for imaging analysis tasks, but also critically erodes clinician and patient trust. Karyotyping is a vital cytogenetic task for detecting genetic abnormalities by analyzing metaphase chromosome images [5]. The process involves first preparing a complete set of microphotographed metaphase chromosomes from cell samples. This includes properly segmenting, classifying, and pairing the 23 chromosome types into homologous pairs to produce a karyogram (see Figure 1). Subsequently, the karyogram is then carefully analyzed by clinical cytogeneticists to identify any anomalies [6][7].