Automated discrimination of dicentric and monocentric chromosomes by machine learning-based image processing

Area of Science:

• Cytogenetics
• Radiation Biology
• Computational Biology

Background:

• Radiation dose estimation relies on analyzing dicentric chromosome (DC) frequencies in peripheral blood lymphocytes.
• Manual analysis of these chromosomal aberrations is time-consuming and subjective.

Purpose of the Study:

• To develop and validate an automated machine learning (ML) approach for detecting dicentric chromosomes (DCs) in Giemsa-stained metaphase cells.
• To assess the accuracy of the automated method in differentiating DCs from monocentric chromosomes (MCs) across varying radiation doses.

Main Methods:

• Image processing techniques including segmentation, chromosome separation (watershed transformation), and feature extraction were employed.
• A Support Vector Machine (SVM) and a Boosting classifier were trained using 14 and 16 image features, respectively, to classify chromosomes.
• The ML model was trained on 292 DCs and 3135 MCs and tested on cells exposed to low (1 Gy) and high (2-4 Gy) radiation doses.

Main Results:

• The automated algorithm demonstrated the ability to differentiate between DCs, MCs, overlapped chromosomes, and debris.
• Performance metrics like True Positive Rate (TPR) and Positive Predictive Value (PPV) were evaluated at different tuning parameters (σ).
• At high doses, TPR ranged from 0.52 to 0.65, with corresponding PPVs from 0.72 to 0.83; at low doses, TPR reached 0.67 with a PPV of 0.26.

Conclusions:

• The developed ML-based automated system provides an accurate and efficient method for dicentric chromosome detection.
• This approach holds significant potential for improving the reliability and throughput of biological dosimetry in radiation exposure assessment.
• The algorithm's performance is acceptable across a range of radiation exposures, offering a valuable tool for cytogenetic analysis.