Figure 3
From: A Perlin Noise-Based Augmentation Strategy for Deep Learning with Small Data Samples of HRCT Images
Examples of 10 ROIs (size of 20 × 20 pixel) of each class of DILD image patterns in HRCT. Examples of the CNN training process: (a) a batch of 16 ROIs generated using Perlin noise, (b) corresponding masks representing the DILD classes for the batch; normal (dark blue), GGO (light blue), consolidation (cyan), RO (yellow), emphysema (orange), and honeycombing (brown), (c) softmax probability maps for the batch; high probability (red), low probability (green), and (d) predicted labels; color scheme is the same as (b). We note that the figure is best viewed in color.
