A pixel-to-pixel segmentation of DILD
without masks
using CNN and Perlin noise
2016.11 njkim@jamonglab.com
Objectives
● Segmenting and labeling regional patterns in
DILD(Diffuse Interstitial Lung Disease) HRCT
images.
From : Younjun Chang et al, “Fast and efficient lung disease classification using hierarchical
one-against-all SVM and cost-sensitive feature selection”. 2012.
Challenges
● Small dataset
○ only 547 ROI ( 20x20 bounding box ) patches
● No human mask label
○ Extremely expensive
Dataset
Dataset
Dataset
Dataset
Traditional approach
● Superpixel approach
Traditional approach
● Superpixel result - factor 0.25
Traditional approach
● Superpixel result - factor 2
Traditional approach
● Superpixel result - factor 4
Traditional approach
● Superpixel result - factor 9
Traditional approach
● Superpixel accuracy
Traditional approach
● Superpixel limitation
○ deterministic and strong assumption
( Similarity of neighboring pixels )
New approach
● Deep learning pixel-to-pixel segmentation.
○ Hand labelled mask is needed.
○ Let’s generate it !
From : Ra Gyoung Yoon et al, “Quantitative assesment of change in regional disease patterns on
serial HRCT of fibrotic interstitial pneumonia with texture-based automated quantification system”.
2012.
Mask generation
● A naive approach → Failed.
○ Because the neural network have learned deterministic
patterns instead of lung disease patterns.
Honeycombing
Emphysema
Mask generation
● Ken Perlin, “An image Synthesizer”, 1985
○ natural appearing textures
○ gradient based fractal noise
○ heavily used in game business
Mask generation
● One random Perlin noise ( simplex noise )
● two randomly selected ROI patches
ConsolidationGGO
Mask ROI Patch
Mask generation
● 547 patches → Infinite patches ( O(1006xN
) )
Model architecture
● UNet + SWWAE architecture
○ Olaf et al, “U-Net: Convolutional Networks for Biomedical Image
Segmentation”, 2015
○ Junbo et al, “Stacked What-Where Auto-encoders”, 2015
Model architecture
Skip connections
Deep learning approach
● pixel-to-pixel segmentation result
Deep learning approach
● pixel-to-pixel segmentation result
Deep learning approach
● pixel-to-pixel segmentation result
Deep learning approach
● pixel-to-pixel segmentation accuracy
High resolution segmentation
● 20 x 20 patches per 512 x 512 image
○ (512 - 20 + 1)2
→ Too expensive
High resolution segmentation
● Fully convolutional layer used
○ Various sized image input available
High resolution segmentation
● 200 x 80 grids
High resolution segmentation
● 500 x 20 grid ( Vertical grids )
High resolution segmentation
● 20 x 500 grid ( Horizontal grids )
High resolution segmentation
● Computation complexity
High resolution segmentation
● Results ( Hortz )
High resolution segmentation
● Results ( Vert )
High resolution segmentation
● Results ( Mix )
High resolution segmentation
● Comparison - Accuracy
High resolution segmentation
● Comparison - computation time
Our contributions
● A simple and practical pixel mask generation
method for DILD ROI dataset using Perlin noise.
○ No radiologist mask needed.
● We applied state-of-the-art deep CNN based
pixel-to-pixel segmentation method to DILD
dataset.
○ High accuracy with reasonable computing time.
Thank you !!!

A pixel to-pixel segmentation method of DILD without masks using CNN and perlin noise