Fully-convolutional neural networks ensemble for comet segmentation in single cell gel electrophoresis assay images

Single-cell gel electrophoresis or comet assay is a simple and reliable method that reveals DNA damage and repair in individual cells. This method is widely used in the medical and biological fields for genotoxicity testing and human biomonitoring studies. Over the past few years, several computational systems have been proposed to perform comet assay image analysis based on traditional image processing methods rather than up-to-date efficient deep learning techniques. Therefore, this work proposes an automatic segmentation system based on a fully-convolutional neural networks ensemble to process comet assay individual cells. Our database was created from peripheral blood samples and consists of comet cells images related to different levels of DNA damage, randomly divided into training, validation and test sets. The experimental results showed that the proposed segmentation model achieves a competitive performance of 87.5% F1-score, 91.5% precision and 84.4% recall. Compared to free-use comet assay analysis systems that apply common thresholding techniques, our model showed more efficient, robust and reliable results. Our model and database are freely available for scientific and academic purposes.