2023 Fiscal Year Final Research Report
Automatic creation of a large amount of virtual normal and abnormal medical images
| Project/Area Number |
21K12722
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 90130:Medical systems-related
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Keywords | 医用画像処理 / 架空画像作成 / 架空病変作成 / 胸部単純写真 / 肺癌 / 病変検出AI |
|---|
| Outline of Final Research Achievements |
In this study, we developed a method to generate synthetic chest X-ray images and further create and naturally embed synthetic nodular lesions. Using interpolation techniques, we achieved even more natural embedding. We generated over 130,000 such synthetic images and used them to train and establish a system for detecting nodules in real chest X-rays. We proposed a new loss function term to effectively utilize the image pairs before and after the nodule embedding and experimentally validated its usefulness. Although the sensitivity of the trained nodule detection system does not necessarily reach the latest state-of-the-art performance, it is quite close to the performance reported in several recent papers, demonstrating the usefulness of this approach.
|
| Free Research Field |
医用画像処理
|
| Academic Significance and Societal Importance of the Research Achievements |
胸部単純写真を対象とした医用診断AIは多数が報告され、一部は実用もされているが、その性能は完ぺきではなく、性能のさらなる向上が求められている。本研究では、そのカギを握る「検出が難しい結節」を無限に自動生成できるアルゴリズムを作成することができ、実際にそれを用いてAIを学習することができた。このアプローチをほかの医用AIにも応用して行くことで、さまざまなAIの性能向上に資せる可能性があり、さらに医師を支えられるAIを開発して行けると信じている。
|
