A multi-pretraining U-Net architecture for semantic segmentation

Pathological cancer research relies heavily on different domain-specific applications including nucleus segmentation from histopathology images. Nucleus segmentation is one of the most challenging tasks because of the many hurdles involved such as masking operations, inaccurate and erroneous annotations, unclear boundaries, poor colours, and overlapping cells. New developments in the deep learning field contributed to the development of new application domains and this has made segmenting nuclei possible. In this research, we propose and evaluate a modified version of a deep learning algorithm called U-Net architecture for partitioning histopathological images. Particularly, we present a novel non-sequential multi-pretraining U-Net architecture and demonstrate that employing a number of persistent parallel models can boost the effectiveness of the segmentation procedures. The proposed approach makes advantage of data augmentation to generate newly synthesized images, which are subsequently processed using a watershed mask. For the validation of the proposed model, we used data from 21,000 cell nuclei at a resolution of 1000 by 1000 pixels. Experimental results demonstrate that the suggested architecture successfully segments nuclei with minimal loss in accuracy.