Scalable model for segmenting Cells' Nuclei using the U-NET architecture

Medical image segmentation significantly influences medicine for diagnostic purposes where higher precision and accuracy are demanded. One of the most cutting-edge architectures in Deep Learning (DL) designed to perform on a few training biomedical images is the U-Shape Network (UNet). Unfortunately, thousands of training biological images are out of reach in most situations, limiting the number of attainable training sets and, as a result, the ability to achieve better accuracy. Therefore, this research proposes a scalable model where the U-Net architecture has been tuned to improve its ability to perform on fewer training samples, mitigating the limited trainable data problem. Several model layers, nodes, hyper-parameters, and normalization methods have been experimented with to tune the original U-Net architecture well, considering possible underfitting and/or overfitting scenarios while maintaining reasonable accuracy. The 2018 Data Science Bowl dataset has experimented with several smaller subsets, 10%-100% of the original dataset, which has been trained separately. The model's accuracy ranged between 96.98% and 97.93% for the experimented subsets. More importantly, the reported variance in the accuracy given different training data sizes is minimal (around 1%).