This document presents a method for unsupervised low-dose CT denoising using an invertible generator in a cycle-free CycleGAN framework. Conventional CycleGAN models require two generators and discriminators, while the proposed method uses a single invertible generator and discriminator. This reduces model complexity while achieving comparable denoising performance to state-of-the-art unsupervised methods. Quantitative and qualitative results on low-dose CT datasets demonstrate the effectiveness of the proposed cycle-free CycleGAN for parameter-efficient unsupervised low-dose CT denoising.
Passive Air Cooling System and Solar Water Heater.ppt
Cycle-free CycleGAN using invertible generator for unsupervised low-dose CT denoising
1. Cycle-free CycleGAN using Invertible Generator for
Unsupervised Low-Dose CT Denoising
가역적 뉴럴 네트워크를 활용한 저선량 전산단층촬영 영상 잡음 제거에 관한 연구
Taesung Kwon and Jong Chul Ye
Taesung Kwon
2. 2
1. Introduction
Cancer related to CT scan, USA, 2007 [2]
High
DNA damage [1]
✓ Potential of increasing the incidence of cancer → Minimizing radiation dose
CT
X-ray
Radiation
dose
Low
✓ Complex noise pattern appears on Low-Dose CT image
Low-dose CT data
Standard-dose CT data
Tube current 20%↓
Lower SNR
Complex noise pattern
Solution
Deep learning
based Low-dose
CT denoising
[1] Huang et al. "DNA damage response signaling pathways and targets for radiotherapy sensitization in cancer." Signal transduction and targeted therapy 5.1 (2020): 1-27.
[2] Berrington et al. “Projected cancer risks from computed tomographic scans performed in the United States in 2007”. Arch Intern Med. 2009 Dec 14;169(22):2071-7.
3. 3
1. Introduction
Uneven distribution of the incident X-ray
photon follows Poisson distribution
Main property of Poisson distribution
• mean = variance = λ
• noise ∝ signal
• SNR(Signal-to-Noise Ratio) ∝
signal
signal
= signal
Detector Plate
Uneven distribution
of X-ray photons
6. 6
2. Background Theory
• Supervised Method
Noisy measurement Paired clean
measurement
• The simplest and fastest method
• Over-blurring problem
• Supervision with paired data -> Hard to get in real case
Output
minimize
7. 7
2. Background Theory
Clean Image Discriminator
Minimize Statistical Distance
Noisy measurement Unpaired clean
measurement
Output
• Unsupervised Method
• No requirement on paired data
• Solve over-blurring problem
8. 8
2. Background Theory
[3] Zhu et al. "Unpaired image-to-image translation using cycle-consistent adversarial networks." Proceedings of the IEEE international conference on
computer vision. 2017
✓ 2 Generators
✓ 2 Discriminators
• To maintain image level information, Cycle-consistency is required.
• Cycle consistency loss
• Requires 2 Generators and 2 Discriminators (Total 4 Neural Networks)
• Unsupervised Method
(CycleGAN [3])
9. 9
2. Background Theory
✓ CycleGAN
LDCT SDCT
LDCT → SDCT Generator
Standard-dose Discriminator
① Forward
SDCT LDCT
SDCT→ LDCT Generator
Low-dose Discriminator
② Backward
✓ Motivation: Cycle-free CycleGAN?
LDCT SDCT
LDCT ↔ SDCT
Invertible Generator
Standard-dose Discriminator
• With Forward Generator & Backward Generator
• Using Invertible Generator automatically fulfills cycle-consistency. Therefore, the backward mapping generator can be eliminated.
10. 10
3. Methods
✓ Neural Network performs non-invertible operations.
Because it uses Non-linear activation function.
Simple neural network calculation
(w is learnable weight, f is activation function) ReLU(Non-linear activation function)
𝑦 = 𝑓 𝑥
𝑓 𝑥 = max(0, 𝑥)
11. 11
3. Methods
Additive coupling layer operation
(Essential component using Neural Network)
✓ Using Coupling layer, invertible calculation is possible.
Designed to enable invertible calculation regardless of Function F. (F can be any neural network)
ℱ1 ℱ1
Inverse additive coupling layer operation
(Essential component using Neural Network)
Identical
12. 12
3. Methods
⋯
✓ By stacking coupling layer with invertible components, we make invertible neural network.
24. 5. Discussion
Comparison with Other CycleGAN Structures
• (a) Conventional cycleGAN: two generators & two discriminators
• (b) Cycle-free CycleGAN: one generator & one discriminator
24
25. 25
5. Discussion
Conventional CycleGAN AdaIN CycleGAN
(Conventional SOTA model)
Proposed
Network # of parameters Network # of parameters Network # of parameters
Total 18,035,202 Total 11,707,843 Total 4,353,409
Table 2. Comparison of the network complexity in terms of trainable parameters
✓ Proposed method uses only 24% of the conventional unsupervised method.
26. 26
5. Discussion
• Proposed method shows parameter-efficient performance
• Also, it shows state-of-the-art denoising performance among comparative methods.
✓ Better performance with reduced complexity
27. 6. Conclusion
• Low-dose CT Denoising
• Minimizing radiation dose, Increasing CT image quality → Important research topic
• Supervised Learning
• Blurry output, Need paired image dataset
• Conventional Unsupervised methods
• No requirement on paired dataset, Need 4 neural networks (heavy computation costs)
• Cycle-free CycleGAN
• Invertible Generator → Use single generator & single discriminator
• Single pair of generator & discriminator → stable training
• Comparable quantitative results with Unsupervised State-of-the-art Methods
• Novel framework for light & stable unsupervised CT denoising
27