Ensembling Unets for Rare Chromosomal Aberration Detection in Metaphase Images, Uncertainty Quantification, and...

Area of Science:

• Biomedical Engineering
• Genetics
• Medical Imaging

Background:

• Biological dosimetry estimates radiation dose using chromosomal aberrations in lymphocytes.
• Manual analysis of metaphase images is time-consuming and prone to errors.
• Current automated systems have high false discovery rates, limiting their utility.

Purpose of the Study:

• To develop a novel deep learning method for automatic detection of rare chromosomal aberrations.
• To improve the reliability and speed of radiation dose estimation in biological dosimetry.
• To quantify uncertainty in automated aberration detection.

Main Methods:

• A deep learning approach formulated as a regression problem with a sparsity-promoting loss.
• Model ensembling using checkpoints from training epochs.
• Consensus voting from artificial experts to provide trustworthy detections and confidence intervals.

Main Results:

• The deep learning method achieved accurate detection of chromosomal aberrations (dicentrics and fragments).
• Radiation dose curves derived from the automated system showed high agreement with reference methods.
• The system provided trustworthy aberration detections with confidence intervals.

Conclusions:

• The proposed deep learning method offers a reliable and efficient alternative for chromosomal aberration detection.
• This advancement can significantly reduce human expertise time and improve diagnosis speed in biological dosimetry.
• The method shows promise for accurate radiation dose estimation and validation.