An End-to-End Chromosome Classification Method Based on Contrastive and Sparse Representation Learning

Chromosome classification is essential in karyotype analysis within the field of cytogenetics. Karyotype analysis is an essential method used by clinical cytogeneticists to identify numerical and structural chromosomal abnormalities. Despite the development of various methods, accurately and robustly classifying chromosomes and their polarity for clinical applications remains a challenging task. These challenges primarily arise from two aspects: 1) the diverse deformations in chromosome shape, 2) the similarity between chromosomes. To tackle these obstacles, we suggest an innovative end-to-end approach in this manuscript. Unlike existing methods, our approach focuses on learning sparse and contrastive features rather than simply identifying the classification features within the given samples. The methodology follows a dual-stage structure, comprising a feature extraction module and two modules dedicated to enhancing features. In the first stage, the basic visual features of chromosome are acquired. In the second stage, the contrastive features are obtained by triplet loss and the sparse features by ista-block. The assessment outcomes on a private dataset indicate that our approach attains 97.38
^%
and 97.76% accuracy for the type task and polarity task, respectively, surpassing state-of-the-art methodologies.