Chromosome Segmentation via Data Simulation and Shape Learning.

Area of Science:

• Cytogenetics
• Computational Biology
• Medical Imaging

Background:

• Karyotyping is crucial for detecting chromosome abnormalities.
• Automatic chromosome classification is advanced, but manual segmentation is a bottleneck.
• Automatic chromosome segmentation faces challenges with limited annotated data and complex region combinations.

Purpose of the Study:

• To develop effective strategies for automatic chromosome segmentation.
• To overcome data scarcity in training segmentation models.
• To accurately identify individual chromosomes from segmented regions.

Main Methods:

• Proposed two data simulation strategies for augmenting training datasets.
• Introduced an optimization-based shape learning method for evaluating chromosome shape.
• Applied methods to a public dataset for validation.

Main Results:

• Data simulation significantly improved segmentation Dice coefficients by 15.8% (non-overlapped) and 46.3% (overlapped regions).
• The optimization-based method achieved 96.2% accuracy in separating overlapped chromosomes.
• Demonstrated the effectiveness of proposed methods on a public dataset.

Conclusions:

• The proposed data simulation and shape learning methods enhance automatic chromosome segmentation.
• These advancements address key challenges in karyotyping and cytogenetic analysis.
• The study contributes to more accurate and efficient chromosome abnormality detection.