I reviewed 3 papers at 'SNU TF Study Group' in Korea.
3 papers tried to solve segmentation problems in medical images with Deep Learning.
Deep Learning 을 이용하여 의료 영상에서 Segmentation 문제를 풀고자 한 3가지 논문을 리뷰하였습니다. :)
The Codex of Business Writing Software for Real-World Solutions 2.pptx
Segmentation problems in medical images
1. 2 4 O c t . 2 0 1 7
Segmentation Problems
in Medical Images
Jimin Lee
Radiological Physics Laboratory,
Seoul National University
SNU TF 스터디 모임
2. Contents
2
1. Spinal cord gray matter segmentation using deep dilated convolutions
( https://arxiv.org/abs/1710.01269 )
2. H-DenseUNet: Hybrid Densely Connected UNet for Liver and Liver Tumor
Segmentation from CT Volumes ( https://arxiv.org/abs/1709.07330 )
3. Automatic Myocardial Segmentation by Using A Deep Learning Network in
Cardiac MRI ( https://arxiv.org/abs/1708.07452 )
3. Spinal cord gray matter segmentation
using deep dilated convolutions
October 4, 2017
* Article under submission to Nature Scientific Reports.
4. 1. Introduction
4
Purpose : To devise an end-to-end fully automated human spinal cord gray matter
segmentation method using Deep Learning
Gray matter (GM) tissue changes in spinal cord (SC)
• It is linked to neurological disorders.
• SCGM atrophy (위축) is a relevant biomarker for predicting disability in amyotrophic
lateral sclerosis (루게릭병).
The fully-automated segmentation is very useful for longitudinal studies.
• The delineation of gray matter is very time-consuming.
5. 1. Introduction
5
Accurate segmentation of the GM is still a remaining challenge.
• Inconsistent surrounding tissue intensities
• Pathology-induced changes in the image contrast
• Differences in MRI acquisition protocols
• Lack of standardized data sets
• Different pixel sizes
• Image artifacts
⋮
MRI samples from different centers → Variability
6. 2. Methods and Materials
6
Dilated convolution
https://github.com/vdumoulin/conv_arithmetic
3x3 Kenel with dilation rate 2
7. 2. Methods and Materials
7
Proposed method
2 layers each
8. 2. Methods and Materials
8
Loss : DSC (Dice Similarity Coefficient)
Data augmentation : rotation, shifting, scaling, flipping, noise, elastic deformation
p : Predictions, r : Gold standard
* 𝜖 term is used to ensure the loss stability by avoiding the numerical issues.
9. 2. Methods and Materials
9
Data sets
(1) Spinal Cord Gray Matter Challenge
• 80 healthy subjects (20 subjects from 4 centers)
• 3 different MRI systems (Philips Achieva, Siemens Trio, Siemens Skyra)
• Training set : 40 subjects, Test set : 40 subjects
10. 2. Methods and Materials
10
Data sets
(2) Ex vivo high-resolution spinal cord
• An entire human spinal cord
• 7T horizontal-bore small animal MRI system
• 4676 axial slices with 100 𝜇𝑚 resolution
• (120 hours to take..!)
11. 2. Methods and Materials
11
Training Protocol (Spinal Cord Gray Matter Challenge dataset)
• Resampling and cropping : 0.25 x 0.25 mm (resampled) / 200 x 200 pixels (cropped)
• Normalization : mean centering, SD normalization
• Train/validation set split : 32 subjects / 8 subjects
• Batch size = 11 (samples), 1000 epochs (32 batches at each epoch)
• Optimization : Adam optimizer (Learning rate : 0.001)
• Batch Normalization : momentum ∅ = 0.01
• Dropout : rate of 0.4
• Learning rate scheduling (𝜂 𝑡0
: initial learning rate, N : the number of epochs, p = 0.9)
16. 4. Discussion
16
They devised a simple, but efficient and end-to-end method.
Compared to U-Net, proposed method provides better results in many metrics.
A major parameter reduction (more than 6 times)
IU : Intersection over Union
18. 1. Introduction
18
Purpose : Liver and liver tumor segmentation in contrast-enhanced 3D abdominal CT scans
Liver and liver tumor segmentation can assist
doctors in making accurate hepatocellular
carcinoma evaluation and treatment planning.
Automatic segmentation is a very challenging task.
• Tumor : Various size, shape, location and
numbers within one patient
• Not clear boundaries
• CT voxel spacing ranges from 0.45mm to
6.0mm
19. 2. Method
19
Proposed method : A novel hybrid densely connected UNet (H-DenseUNet)
U-Net
U-Net: Convolutional Networks for Biomedical Image Segmentation ( https://arxiv.org/abs/1505.04597 )
To aggregate
semantic information
To recover the spatial
information with the help of
shortcut connections
23. 3. Experiment and Results
23
Data set
• MICCAI 2017 LiTS (Liver Tumor Segmentation) Challenge
• Contrast-enhanced 3D abdominal CT scans
• 131 for training, 70 for testing
Evaluation Metrics
• Dice per case score : An average Dice per volume score
• Dice global score : Dice score evaluated by combining all datasets into one