Auto-encoders for Detection and Counting of Live/Dead Cells

Counting cells and detecting their viability is a task of utmost importance in the medical field, but this task has only 2 ways, namely manual and automatic cell counting. Manual cell counting is reliant on the human counting the cells and classifying them as viable or non-viable and therefore is prone to human error is monotonous and time-consuming and relies on the experience of the counter. Automatic cell counting is accurate, but not all labs are equipped with an automatic counter, and that is because of its high price. One could leverage the power of machine learning to create a machine that could count cells and classify them effectively and without the high cost of automatic cell counters. The methodology contains the methods of acquiring a dataset, how to create binary images for this dataset, and experimentation with 3 different machine learning models, namely VGG-16 Unet, ResUnet, and Attention ResUnet. The models predict maps of the following accuracy values: VGG-16 U-net scored 82%, ResUnet scored 86%, and Attention ResUnet 82%, with training images equal to 402 images, which is considered very good compared to other models that usually use data sets with double or triple the images used in this paper.