Explainable Deep Learning Framework for Automated Human Chromosome Classification

Abstract

Abstract—Most laboratories continue to perform karyotyping manually because this process needs trained cytogeneticists to identify all 24 human chromosome classes from G-banded metaphase images. The process remains slow because it depends on expert availability for its completion. The application of deep learning models for chromosome classification in medical practice faces multiple barriers because of restricted public data availability and uninterpretable results and complex decision mechanisms of CNNs. Our research introduces an explainable deep learning system which delivers both precise results and clear understanding of its operations. The CNN-based models received training from Cell Image Library G-banded chromosome images while explanation methods Grad-CAM and LRP and Integrated Gradients generated visualizations of model prediction processes. The system evaluation combined numerical performance metrics with expert evaluation of explanation effectiveness. The model achieved 98% accuracy on test data while explanation maps demonstrated the presence of essential chromosome characteristics including banding patterns and centromere positions and distinctive shapes. The model attention patterns followed cytogenetic evaluation standards but it failed to distinguish between chromosomes with similar appearances. The research indicates that explainable deep learning systems can perform automated karyotyping through precise results and detailed visual explanations which enhance medical trust in clinical practice. Keywords—Explainable Artificial Intelligence (XAI); Chromosome classification; Deep learning; Convolutional Neural Networks (CNNs); Grad-CAM; Layer-wise Relevance Propagation (LRP); Integrated Gradients; Cytogenetics; Karyotyping automation; Biomedical image analysis; Medical diagnostics; Image-based classification; Interpretable machine learning; Computer vision in healthcare.

Related publications

• Convolutional Neural Networks with Transfer Learning for Three-Class Lung CT ClassificationIEEE International Conference on Smart Information Systems and Technologies · 2026 · Same author: Alisher Dautov
• XXIV international scientific conference. London. Great Britain. 04-05.12.2025Zenodo · 2025 · Same author: Alisher Dautov
• Enhancing Cardiovascular Disease Classification with Routine Blood Tests Using an Explainable AI ApproachAlgorithms · 2025 · Citations: 2 · Shared topic: Explainable Artificial Intelligence (XAI)
• Beyond Attention: Module-Level Refusal Directions in Mamba-Based Hybrid Language ModelsLecture notes in computer science · 2026 · Shared topic: Explainable Artificial Intelligence (XAI)