| Project/Area Number |
23K11170
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 61010:Perceptual information processing-related
|
|---|
| Research Institution | Hiroshima University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2023-04-01 – 2026-03-31
|
| Project Status |
Granted (Fiscal Year 2023)
|
| Budget Amount *help |
¥4,420,000 (Direct Cost: ¥3,400,000、Indirect Cost: ¥1,020,000)
Fiscal Year 2025: ¥1,170,000 (Direct Cost: ¥900,000、Indirect Cost: ¥270,000)
Fiscal Year 2024: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
Fiscal Year 2023: ¥1,690,000 (Direct Cost: ¥1,300,000、Indirect Cost: ¥390,000)
|
|---|
| Keywords | deep learning / automatic image analysis / uncertainty / neural networks / image analysis / reliable AI |
|---|
| Outline of Research at the Start |
In this research project novel deep learning based machine learning methods will be investigated for automatic bio-medical image analysis, which can work reliably in critical settings where safety is first priority.
|
| Outline of Annual Research Achievements |
In recent years deep neural networks have achieved SOTA(state-of-the-art) performance on many automatic image analysis tasks, even surpassing human-level performance in certain cases. However, they still perform less than satisfactory when test data differs significantly from the data with which they have been trained with (and especially in cases where the training data sets are small). This is particularly important in bio-medical images, where over-confidence in the predictions can lead to life-threatening situations and in images obtained when studying physics-related phenomena, which differ significantly from the natural images used to train the current SOTA deep learning methods.
During the 1st year of the project we have experimented with using Bayesian neural networks to quantify the uncertainty and a baby-step curriculums learning method which mimics how people learn starting from easy data and gradually moving to difficult one. In order to obtain a more objective evaluation of the efficacy of the proposed methods we have decided to use more difficult data: e.g. skin lesions data which is widely used for many medical challenge competitions for medical images, and images of bag-breakup-related fluid phenomena for more general scientific images. In the latter case, we developed a novel algorithm which utilizes information from several consecutive frames and significantly improves accuracy of detection (even compared with present SOTA algorithms) by eliminating false positives which are difficult to discriminate from a single image even for human experts.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The initial plan was to investigate novel machine learning methods (like for example curriculum learning based methods for quantifying and reducing the uncertainty in the automatic analysis of bio-medical images (one concrete example being images of iPS cell colonies, with which we had done some pioneer work before) in order to contribute to the design of more reliable automatic image analysis systems which can work reliably in the presence of uncertainty, especially in cases where it is difficult to obtain a huge number of training images. Obtained experimental results during the first year of the project have shown that better results can be obtained by not limiting our attention to curriculum learning, but for example combining it with novel data augmentation methods (for bio-medical images), or using temporal information from several consecutive frames (in the case of general images of physical phenomena like bag-breakups).
|
| Strategy for Future Research Activity |
During the next stage of the project we plan to continue the development of novel deep learning methods for improving the accuracy and reducing the uncertainty in the automatic analysis of both general and bio-medical images, thus endeavoring to contribute to the development of more reliable and trustworthy AI-based systems. We plan to continue our effort for developing a curriculum learning method combined with novel data augmentation methods in the context of bio-medical images and additionally to experiment with some other types of data where uncertainty can lead to life-threatening situations, like for example with data obtained from the cameras of self-driving vehicles.
|
