2021 Fiscal Year Final Research Report
Development of noise removal method using unsupervised learning for SPECT images
Project/Area Number |
20K20186
|
Research Category |
Grant-in-Aid for Early-Career Scientists
|
Allocation Type | Multi-year Fund |
Review Section |
Basic Section 90110:Biomedical engineering-related
|
Research Institution | Fujita Health University |
Principal Investigator |
Shirakawa Seiji 藤田医科大学, 保健学研究科, 准教授 (50308847)
|
Project Period (FY) |
2020-04-01 – 2022-03-31
|
Keywords | ノイズ除去 / SPECT / Deep Learning / 核医学 |
Outline of Final Research Achievements |
The SPECT (Single Photon Emission CT) image of the nuclear medicine examination contains a lot of noise because the count number is small, and the image becomes an image. Therefore, it is necessary to remove noise before using it for diagnosis. This study developed a noise removal method by deep learning for myocardial blood flow SPECT images. In this method, the network was optimized by unsupervised learning for nuclear medicine images for which ideal (noise-free) images could not be obtained, and noise was removed. As a result, it has become possible to remove noise while maintaining spatial resolution, compared to the conventional Gaussian filter processing.
|
Free Research Field |
核医学
|
Academic Significance and Societal Importance of the Research Achievements |
核医学、特にSPECT画像はノイズを多く含んだ画像であり、ノイズ除去処理は必須である。現在、使用されているGaussianフィルタなどの処理は、検査ごとにパラメータを設定する必要があり、処理者や施設ごとに結果が異なる恐れがある。またGaussianフィルタなどはノイズ除去効果が高くなるほど空間分解能が低下する問題がある。 我々が開発したDeep Learningを用いたノイズ除去法は、パラメータ設定を必要とせず、画像のノイズ成分を抽出し、原画像から減算するため空間分解能劣化が生じない。さらにGaussianフィルタと同等のノイズ除去効果を有し、臨床的意義は高いと示唆される。
|