DnCNN을 이용한 X선 흉부 영상의 잡음 제거

다국어 초록 (Multilingual Abstract)

Radiation exists everywhere around us. Radiation existing in nature is called natural radiation, and radiation made for human use is called artificial radiation. The majority of artificial radiation is medical radiation such as X-ray and CT. Radiation causes damage to DNA, transformation of cells and is known to cause disease or tumors. X-rays can also accumulate and affect tumor induction.

In the chest x-ray image, you may have an unwanted pixel value called Quantum noise that degrades image quality for various reasons such as electrical interference and environmental factors. The noise can be reduced by increasing the dose of X-ray for production of X-ray image but it will harm patient body.

In this paper, the DnCNN model based on deep learning which removes additive white Gaussian noise is proposed and its performance is compared with Wiener filter. As a result, MAE value obtained by DnCNN from the chest x-ray image was lower than that by Wiener filter. Therefore, the DnCNN reached better performance in image denoising compared with Wiener filter.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• Ⅱ. 이론적 배경 6
• 1. 머신러닝(Machine Learning) 6
• 2. 딥러닝(Deep Learning) 9
• 3. CNN(Convolutional Neural Networks) 11
• (1) CNN의 배경 11
• (2) CNN의 구조 12
• (3) CNN의 특징 13
• 4. 위너필터(Wiener Filter) 15
• (1) 위너필터의 개념 15
• (2) 위너필터의 구조 15
• 5. DnCNN(Denoising Convolutional Neural Network) 20
• (1) DnCNN의 개념 20
• (2) DnCNN의 구조 20
• (3) DnCNN의 잡음 제거원리 23
• Ⅲ. 연구 방법 25
• 1. 대상 및 분석 도구 25
• 2. 비교 및 분석 절차 26
• Ⅳ. 연구 결과 27
• 1. Read Dicom 영상 27
• 2. 3056x2517 16bit → 1024x1024 16bit ROI 28
• 3. 1024x1024 16bit → 512x512 8bit Down-sampling 28
• 4. 원본 vs. Noisy vs. DnCNN 29
• 5. 원본 vs. DnCNN vs. 위너필터 31
• 6. 512x512 8bit → 1024x1024 16bit Up-sampling 33
• Ⅴ. 결론 및 고찰 34