Deep neural networks show high accuracy in the problem of semantic and instance segmentation of biomedical data. However, this approach is computationally expensive. The computational cost may be reduced with network simplification after training or choosing the proper architecture, which provides segmentation with less accuracy but does it much faster. In the present study, we analyzed the accuracy and performance of UNet and ENet architectures for the problem of semantic image segmentation. In addition, we investigated the ENet architecture by replacing of some convolution layers with box-convolution layers. The analysis performed on the original dataset consisted of histology slices with mast cells. These cells provide a region for segmentation with different types of borders, which vary from clearly visible to ragged. ENet was less accurate than UNet by only about 1-2%, but ENet performance was 8-15 times faster than UNet one.

No abstract