SupCAM: Chromosome cluster types identification using supervised contrastive learning with category-variant...

Area of Science:

• Genetics and Genomics
• Computational Biology
• Medical Imaging

Background:

• Chromosome segmentation is vital for karyotyping and identifying chromosomal abnormalities.
• Existing methods struggle with touching and occluding chromosomes, necessitating improved cluster type identification.
• Previous cluster identification methods relied on large natural image datasets, overlooking semantic differences with chromosomes.

Purpose of the Study:

• To develop a novel method, SupCAM, for accurate chromosome cluster type identification.
• To overcome limitations of small-scale datasets by avoiding reliance on natural image datasets.
• To enhance automatic chromosome segmentation through improved cluster classification.

Main Methods:

• A two-step approach: supervised contrastive pre-training on ChrCluster, followed by fine-tuning.
• Introduced category-variant image composition for data augmentation.
• Incorporated self-margin loss with angular margin to improve intra-class consistency and decrease inter-class similarity.

Main Results:

• SupCAM achieved 94.99% accuracy on the ChrCluster dataset, outperforming previous methods.
• Effectiveness of individual modules was validated through extensive ablation studies.
• The method successfully avoided overfitting, demonstrating robust performance solely on the ChrCluster dataset.

Conclusions:

• SupCAM offers a significant advancement in chromosome cluster type identification.
• The developed method enhances the accuracy of automatic chromosome segmentation.
• SupCAM's approach is effective for analyzing chromosome images with touching and occluding clusters.