- Keywords: Histopathology, Segmentation, Context, Dilated
- TL;DR: We propose a model that aggregates context through use of parallel dilated and standard convolutions embedded within encoder-decoder block of the segmentation network.
- Abstract: Digital histopathology has seen a significant improvement towards aiding pathologists diagnosis by the use of deep learning segmentation models. The segmentation task allows for segregating benign and malignant cells from digitized tissue whole slide images (WSI). However, for reliable deployment, it is essential to produce accurate object boundaries. Pathologists performs this challenging task using context to extract complex features from a WSI. We leverage this idea and propose a model that aggregates context through use of parallel dilated and standard convolutions, embedded within each encoder-decoder block of the segmentation network. We show that, the proposed architecture not only beats the baseline by a large margin but also achieves an accuracy improvement over state-of-the-art segmentation architecture by upto 1.95% DICE Score.