Deep Neural Network-Based Automatic Dicentric Chromosome Detection Using a Model Pretrained on Common Objects.

Area of Science:

• Cytogenetics
• Radiological protection
• Artificial Intelligence

Background:

• Dicentric chromosome assay (DCA) is a key method for estimating radiation dose.
• Current DCA methods are labor-intensive and require specialized skills.
• Automating DCA can improve efficiency and accuracy in biodosimetry.

Purpose of the Study:

• To develop and evaluate a deep learning model for automating the detection of dicentric chromosomes in metaphase images.
• To assess the performance of the YOLOv5 algorithm for dicentric chromosome identification.
• To demonstrate the feasibility of using pretrained models for efficient training with limited data.

Main Methods:

• Utilized YOLOv5, a one-stage object detection algorithm, for automated analysis of chromosome metaphase images.
• Trained the YOLOv5 model on 887 augmented chromosome images, leveraging pretrained parameters.
• Evaluated model performance using validation (380 images) and test (300 images) datasets.

Main Results:

• The pretrained YOLOv5 model achieved a maximum F1 score of 0.94 and a mean average precision (mAP) of 0.961 in detecting dicentric chromosomes.
• A randomly initialized model showed decreased performance, with a maximum F1 score of 0.82 and mAP of 0.873%.
• Results confirm the model's effectiveness in accurately detecting dicentric chromosomes.

Conclusions:

• Deep learning-based object detection using YOLOv5 can effectively automate the dicentric chromosome assay.
• Pretrained models enable efficient training with relatively small datasets for radiation biodosimetry.
• Automated DCA holds promise for faster and more accessible radiation dose assessment.