ChromosomeNet: Deep Learning-Based Automated Chromosome Detection in Metaphase Cell Images

Area of Science:

• Genetics and Genomics
• Computational Biology
• Medical Imaging Analysis

Background:

• Chromosomal disorders arise from abnormal chromosome number or structure, necessitating accurate chromosome screening in prenatal care.
• Manual chromosome analysis is time-consuming and resource-intensive, particularly with the growing demand for prenatal diagnosis.
• An automated approach for chromosome detection and recognition is crucial to alleviate the strain on human labor resources.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for the automatic detection and recognition of chromosomes in metaphase cell images.
• To create a system applicable in clinical settings by processing original images without preprocessing.
• To assess the system's performance, especially on challenging images identified by physicians.

Main Methods:

• A deep learning system, ChromosomesNet, was developed using a large database of 5,000 metaphase cell images (229,852 chromosomes).
• ChromosomesNet integrates one-stage and two-stage model advantages, accepting original images as direct input.
• The system was validated using 3,827 simple and 1,173 difficult images, as classified by medical professionals.

Main Results:

• ChromosomesNet achieved a high accuracy of 99.60% using the COCOAPI's mAP50 evaluation method.
• The system demonstrated superior performance with recall at 99.9% and an F1 score of 99.49%, outperforming five other object detection methods.
• Notably, ChromosomesNet achieved 99.5% accuracy in detecting difficult chromosome images, surpassing previous studies.

Conclusions:

• The proposed ChromosomesNet system exhibits robust stability, strong performance, and practical applicability for clinical use in chromosome analysis.
• The system's effectiveness was validated on a large dataset, including challenging images, exceeding the scope of prior research.
• Further cross-hospital validation with data from diverse clinical settings is recommended to confirm widespread clinical utility.