Chromosome Detection in Metaphase Cell Images Using Morphological Priors

Area of Science:

• Computational biology
• Medical imaging analysis
• Genetics

Background:

• Accurate chromosome detection in metaphase cell (MC) images is crucial for karyotype analysis and diagnosing chromosomal disorders.
• Challenges include dense distributions, varied orientations, and diverse morphologies of chromosomes.
• Existing methods struggle with the complexity of MC images.

Purpose of the Study:

• To propose DeepCHM, a novel rotated-anchor-based detection framework for fast and accurate chromosome detection in MC images.
• To enhance feature representation and optimize anchor selection for improved detection performance.
• To develop a robust method for addressing challenges in chromosome detection.

Main Methods:

• Developed a DeepCHM framework incorporating a deep saliency map for enhanced feature learning and guided anchor setting.
• Implemented a hardness-aware loss function to prioritize difficult chromosome instances.
• Utilized a model-driven sampling strategy to mitigate anchor imbalance during training.
• Created a large-scale benchmark dataset (624 images, 27,763 instances) for training and evaluation.

Main Results:

• The DeepCHM framework achieved a high average precision (AP) score of 93.53%.
• Demonstrated superior performance compared to most state-of-the-art (SOTA) chromosome detection methods.
• Effectively handled dense distributions, arbitrary orientations, and varied morphologies of chromosomes.
• The saliency map, hardness-aware loss, and sampling strategy significantly improved detection accuracy and speed.

Conclusions:

• DeepCHM offers a significant advancement in automated chromosome detection for cytogenetic analysis.
• The proposed innovations effectively address key challenges in chromosome detection from MC images.
• The developed benchmark dataset facilitates further research and development in the field.