KLI

딥러닝 알고리즘을 이용한 영상복원을 통한 저선량 디지털 유방촬영술 플랫폼 개발

Metadata Downloads
Abstract
Purpose:
First, we analyze qualitatively the detection and characterization of breast cancer on low dose mammographic images and find the acceptable low dose levels compared to the full dose image. Second, we use the convolutional neural network (CNN) de-noising method to convert the acceptable low dose images and find the lowest dose to reconstruct the acceptable image quality compared to synthesized full dose image.
Materials and methods:
This prospective study was approved by institutional review board. The true full dose level was determined through AEC, and images were subsequently acquired at 5 different radiation dose levels (80% of AEC, 60%, 40%, 20%, and 10%) of breast cancer mastectomy specimen with digital mammography. For the first purpose, five radiologists evaluated low dose images by comparison to the reference true full dose image and scaled as equivalent, acceptable or unacceptable for each dose level. For the second purpose, the same five radiologists blindly rated three images (low dose, synthesized full dose, and true full dose). We analyzed the trend using Mantel-Haenszel statistic. In addition, we compared the quantitative assessment using McNemar’s or marginal homogeneity test.
Results:
Mass and calcification detection rate decreased substantially at 10% reduced dose. Regarding characterization, the ‘not acceptable’ rate of mass and calcification increased at 10% (86.3%) and 20% (83.8%) respectively. The 80% and 60% images were equivalent to full dose images regarding both mass and calcification detection and characterization. The synthetized images showed high detection rate of 87.4-90.0% and 96.8-100.0% for 20% and 40% respectively. There was significantly higher ‘equivalent’ image quality for both mass and calcification at synthesized 40% dose level (65.3% and 65.0% respectively) compared to synthesized 20% dose level (41.1% and 20.8% respectively) (p<0.001). Compared to low dose images, synthesized image quality improved in 40% dose whereas at 20% dose, it did not increase in synthesized images (p<0.001).
Conclusion:
There is a potential for modest dose reduction retaining diagnostic information with de-noising algorithm in digital mammography. Our results provide a baseline for future studies on reducing the radiation dose with lesion preservation on mammography.
Author(s)
하수민
Issued Date
2018
Awarded Date
2019-02
Type
Dissertation
URI
https://oak.ulsan.ac.kr/handle/2021.oak/6669
http://ulsan.dcollection.net/common/orgView/200000173630
Affiliation
울산대학교
Department
일반대학원 의학과
Advisor
김학희
Degree
Doctor
Publisher
울산대학교 일반대학원 의학과
Language
eng
Rights
울산대학교 논문은 저작권에 의해 보호받습니다.
Appears in Collections:
Medicine > 2. Theses (Ph.D)
Authorize & License
  • Authorize공개
Files in This Item:

Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.