Highly Performing Automatic Detection of Structural Chromosomal Abnormalities Using Siamese Architecture.

Area of Science:

• Genetics
• Bioinformatics
• Artificial Intelligence

Background:

• Structural chromosomal abnormalities (SCAs) are vital for diagnosing genetic diseases and cancers.
• Current detection methods by experts are time-consuming and labor-intensive.

Purpose of the Study:

• To develop an intelligent and high-performing AI method to assist cytogeneticists in screening for SCAs.
• To leverage Siamese Convolutional Neural Networks (CNNs) for detecting chromosomal abnormalities.

Main Methods:

• Utilized Siamese CNN architecture to evaluate similarities between paired chromosomes for abnormality detection.
• Experimented with seven popular CNN models, including Xception and InceptionResNetV2, on a dataset of chromosome 5 deletions (del(5q)).
• Applied data augmentation techniques to enhance model performance.

Main Results:

• Achieved high F1-scores for del(5q) detection, with Xception (97.50%) and InceptionResNetV2 (97.01%) showing superior performance.
• Successfully detected complex inversions (inv(3)), achieving a 94.82% F1-score after specific training.
• Demonstrated the method's effectiveness on multiple types of SCAs.

Conclusions:

• The proposed Siamese CNN-based method is the first highly performing technique for automated SCA detection.
• This AI-driven approach significantly assists cytogeneticists, offering a faster and more accurate screening tool.
• The publicly available code facilitates further research and application in clinical settings.