repository
Special Graduate Schools
Graduate School of Science and Technology
Department of Industrial Engineering
3. Theses(Master)
MRI 영상에서의 유방 조직 다중 분할을 위한 Haar Wavelet Pooling기반 컨볼루션 신경망
- Alternative Title
- Haar Wavelet Pooling-based Convolutional Neural Network for Multiple Segmentation of Breast Tissues from MRI Images
- Author(s)
- 양광빈
- Alternative Author(s)
- Yang Kwangbin
- Advisor
- 양정삼
- Department
- 일반대학원 산업공학과
- Publisher
- The Graduate School, Ajou University
- Publication Year
- 2022-08
- Language
- kor
- Keyword
- 다중 클래스 의미론적 분할; 웨이블릿 풀링 컨볼루션 신경; 유방 재건; 이산 웨이블릿 변환
- Abstract
- 본 논문에서는 Haar wavelet pooling 기반의 U-net을 사용하여 서 있는 자세의 유방 형상 구축을 위한 유방 조직들의 Multi-class semantic segmentation을 제안한다. 우선, 유방 절제술 이전에 촬영된 MRI영상으로부터 유방 형상 구축 시 필요한 유방 조직인 피부, 지방, 유선 조직(Fibroglandular tissue) 그리고 배경을 레이블링(labeling)하여 학습 데이터 세트를 구축한다. 이후 유방 조직에 대한 분할 정확도를 개선하기 위해서 Haar wavelet pooling기반 U-net을 이용한 Multi-class semantic segmentation을 수행한다. Haar wavelet pooling기반 U-net는 입력 이미지의 근사값을 보유하는 Low-Low 대역(sub-band)과 세부적인 Edge 특징의 정보를 가지는 3개의 distinct frequency sub-bands (Low-High, High-Low, High-High)를 동시에 사용하여 subsampling단계에서 이미지의 정보 손실량을 줄이고 효과적인 유방 조직의 특징을 추출할 수 있다. 또한, 과적합에 강건한 딥러닝 네트워크를 구현할 수 있다. 본 논문에서 제시된 Haar wavelet pooling기반 U-net을 이용해 유방 조직들에 대한 분할 실험을 수행하여 87.48의 mIOU을 얻었다.
- Alternative Abstract
- This study proposes multi-class semantic segmentation of breast tis-sues for reconstructing the breast shape in the standing position using U-net based on Haar wavelet pooling. First, a training dataset is con-structed by labeling the skin, fat, fibroglandular tissue, and the back-ground, which are required for reconstructing the breast shape, from MRI images before mastectomy. Next, multi-class semantic segmentation is performed using U-net based on Haar wavelet pooling to improve the segmentation accuracy for breast tissues. The U-net based on Haar wavelet pooling can simultaneously use a low-low band that holds the approximation of the input image and three distinct frequency sub-bands (low-high, high-low, high-high) with information of detailed edge fea-tures to reduce the loss of the breast information in the subsampling phase. Moreover, a deep learning network robust to overfitting was im-plemented. An mIoU of 87.48, was obtained in segmentation experiments conducted for breast tissues using the proposed U-Net based on Haar wavelet pooling.
- Fulltext
- Appears in Collections:
- Special Graduate Schools > Graduate School of Science and Technology > Department of Industrial Engineering > 3. Theses(Master)
- Files in This Item:
- There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.
-
- License
- STATISTICS
- Total Visit :4,068,364
- Total Download :1,966
- Today View :7,685
