2016 Fiscal Year Research-status Report
Automatic Detection of Good/Bad Colonies of iPS Cells Using Deep Learning
| Project/Area Number |
16K00394
|
|---|
| Research Institution | Hiroshima University |
|---|
Principal Investigator |
ライチェフ ビセル 広島大学, 工学研究院, 助教 (00531922)
|
|---|
| Project Period (FY) |
2016-04-01 – 2019-03-31
|
| Keywords | iPS cells / deep learning / segmentation / CNN / colony detection |
|---|
| Outline of Annual Research Achievements |
Induced pluripotent stem (iPS) cells have shown a huge potential to revolutionize medical therapy by personalizing regenerative medicine and creating novel disease models for research and therapeutic testing. However, in order for this to happen, a steady supply of iPS cells obtained through harvesting of individual cell colonies is needed. The purpose of this research project is the design of a machine learning method for automatic detection of Good/ Bad colonies of iPS cells, which would make possible to automate the cell harvesting process.
Recently deep learning methods, which automatically extract hierarchical features capturing complex nonlinear relationships in the data, have managed to successfully replace most task-specific hand-crafted features, resulting in a significant improvement in performance on a variety of biomedical image analysis tasks. Currently Convolutional Neural Network (CNN) based methods define the state-of-the-art in this area and for this reason in this research our task is to develop a CNN-based method for the automatic detection of Good/Bad colonies of iPS cells. Preliminary experimental results seem to indicate that very good accuracy of detection can be achieved by this approach.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
At this stage we have managed to develop and test to some extent the deep-learning based algorithm for automatic detection of Good/Bad colonies of iPS cells. To achieve maximal precision of detection, rather than using image-level recognition of the colonies, we concentrate on the harder problem of image segmentation, where each pixel needs to be classified into its corresponding class/category. Our approach is to extract local patches from the images (where several different colonies can appear) and to use the CNN as a pixel-wise classifier. During training, the patch is used as an input to the network and it is assigned as a label the class of the pixel at the center of the patch (available from ground-truth data provided by a human expert). During the test phase, a patch is fed into the trained net and the output layer of the net provides the probabilities for each class. We have developed a novel algorithm for patch classification which utilizes structural information to achieve significantly higher precision than previous methods on our data set.
|
| Strategy for Future Research Activity |
More extensive testing needs to be done using different parameters for the deep neural network used in the proposed method to validate the results. Also, alternative methods for combining the information at different resolution levels of the local patches and networks of alternative design need to be studied. We expect that, in this way, the accuracy of detection can be further improved.
|
| Causes of Carryover |
At this stage of the project we concentrated more on the theoretical development of the method for automatic detection of Good/Bad colonies of iPS cells, while the planned use of more powerful computational resources will be necessary for the next stage of the project.
|
| Expenditure Plan for Carryover Budget |
The incurring amount will be used for providing the computational resources to support the project at the next stage.
|
-
-
-
-
-
[Presentation] 深層学習を用いた大腸NBI内視鏡画像認識2016
Author(s)
田中 孝二郎, Bisser Raytchev, 玉木 徹, 小出 哲士, 吉田 成人, 三重野 寛, 田中 信治
Organizer
第19回画像の認識・理解シンポジウム(MIRU2016)
Place of Presentation
アクトシティ浜松, 静岡
Year and Date
2016-08-01 – 2016-08-04
-
