Table 1 Summary of difficulties in segmenting high-resolution images.

From: Understanding important features of deep learning models for segmentation of high-resolution transmission electron microscopy images

Problem Modification Motivation
Particle features not recognized Additional convolutional layer at each up-sampling step Successive convolutional layers work to relate spatial and feature dimensions, not just concatenate29,30
Unclear boundaries at particle edges Preprocessing with Gaussian blur Noise makes actual features difficult to detect, reducing frequency of variation makes physical features apparent
No activation in particle output layer Apply leaky ReLU activation Small activations are pushed towards 0 (“dying ReLU problem”)10,31
Large variation in raw activation values Add batch normalization after each convolutional layer Regularize model variance at each step to maintain original intensity distribution32
  1. Each potential problem is described, along with a data-driven approach to a solution.