Dual attention and multi-scale pyramid network (DAMP): a novel approach to chromosome classification

Karyotyping is a standard procedure in cytogenetics for the assessment of chromosomal abnormalities associated with genetic disorders, cancers, and other diseases. The procedure involves staining and imaging metaphase chromosomes to evaluate chromosome number and structural composition, enabling the detection of anomalies such as aneuploidies, translocations, and structural rearrangements. Manual interpretation of karyograms is time-consuming and subject to inter-observer variability, motivating the development of automated approaches. A classification method, the Dual Attention Multiscale Pyramid Network (DAMP), is presented for chromosome classification into 24 classes. The architecture incorporates a Channel Attention Mechanism (CAM) and a Spatial Attention Mechanism (SAM) within feature-extraction layers to refine representation. Channel-wise feature responses are adaptively recalibrated by CAM through inter-channel dependency modeling, while SAM generates spatial attention maps to emphasize informative regions. A Multiscale Feature Pyramid Network (FPN) is used to capture hierarchical representations across resolutions, improving discrimination of chromosomes with varying sizes, orientations, and structural features. A large-scale dataset comprising metaphase images, karyograms, and 523,074 per-chromosome crops from 1,311 patients is introduced; image-level abnormality tags are provided as metadata for stratified analysis. Unlike prior genotype-centric resources, this work provides an image-level chromosome dataset with abnormality-tag metadata and a domain-adapted dual-attention network. Experimental evaluation indicates that DAMP attains 96.76% classification accuracy on the presented dataset, exceeding the performance of baseline methods. The proposed approach is expected to support reproducible development of automated karyotyping tools.