Non-alcoholic steatohepatitis (NASH) is a prevalent liver disease caused by the accumulation of fat-lipid droplets in the liver cell resulting in health complications. This study focused on training of three deep convolutional architectures for semantic segmentation, namely DUCK-Net, U-Net and U-Net++ using histology images for hepatic steatosis quantification. The performance metrics were evaluated on the trained models using multiple independent image datasets available publicly. The model’s architecture was defined from the ground up and specific attention was paid to different degree of usefulness of all the three architectures. For the purpose of hepatic steatosis liver tissue quantification, the dataset (350 liver histology images) was split into training set of 280 images, with validation set containing 70 images. The trained network was tested with 35 images of the same dataset. Same methodology was adopted for all three networks and metrics were evaluated from the predicted segmentation masks. The trained network was also used to perform the steatosis grading (Testing) for another independent set of 477 images and using these predictions, the steatosis percentage was computed. Using the percentage obtained, the tissue cells were graded for the steatosis severity and these grades were compared against the ground truth grades. These findings indicate the efficacy of the DUCK-Net and other segmentation models for steatosis quantification and grading.