| Project/Area Number |
17K19402
|
|---|
| Research Category |
Grant-in-Aid for Challenging Research (Exploratory)
|
| Allocation Type | Multi-year Fund |
|---|
| Research Field |
Biology of Cells to Organisms, and related fields
|
|---|
| Research Institution | Kyushu University |
|---|
Principal Investigator |
Uchida Seiichi 九州大学, システム情報科学研究院, 教授 (70315125)
|
|---|
| Project Period (FY) |
2017-06-30 – 2019-03-31
|
| Project Status |
Completed (Fiscal Year 2018)
|
| Budget Amount *help |
¥6,240,000 (Direct Cost: ¥4,800,000、Indirect Cost: ¥1,440,000)
Fiscal Year 2018: ¥2,340,000 (Direct Cost: ¥1,800,000、Indirect Cost: ¥540,000)
Fiscal Year 2017: ¥3,900,000 (Direct Cost: ¥3,000,000、Indirect Cost: ¥900,000)
|
|---|
| Keywords | バイオイメージインフォマティクス / 深層学習 / 画像変換 / バイオイメージ・インフォマティクス / 機械学習 / 画像情報学 |
|---|
| Outline of Final Research Achievements |
In this research, we tried to realize universal image processing framework based on machine-learning for bioimage-informatics. Bioimage-informatics is a interdisciplinary research area between image-informatics and biology. Its research topics include image filtering, denoising, segmentation, etc. Main difficulties of bioimage-informatics are 1) its complicated process pipeline and 2) lack of enough training samples. About 1), we need to combine several image processing units to achieve the expected results. To solve those difficulties, we have developed a new machine-learning method, called modular u-net. The idea of modular u-net is to concatenate u-nets, each of which performs a specific image processing, such as binarization, by a neural network-based mechanism. Since each u-net can be pre-trained sufficiently by general images, we can realize the expected image processing quality by a fine-tuning step after the concatenation with a limited number of training samples.
|
| Academic Significance and Societal Importance of the Research Achievements |
我々が開発したModular U-netは，様々な画像処理をニューラルネットワークであるu-netをモジュールとして実現するという新たな枠組みであり，生体画像以外にも様々な用途に利用しうる．各u-netは一般的な画像を使って十分に事前学習しておけるために，最終的な応用先における学習サンプルが希少であっても実用化のなところが強みである．実際，我々はすでに古文書画像処理の分野でmodular u-netの有効性を定量的に示している．
|
