2022 Fiscal Year Final Research Report
Computer-aided detection development based on semi-supervised learning with medical chart information
| Project/Area Number |
20K11944
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 61030:Intelligent informatics-related
|
|---|
| Research Institution | Kindai University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | コンピュータ支援診断AI / 弱教師学習 / 医用画像 / 3D U-Net / 適応的閾値処理 |
|---|
| Outline of Final Research Achievements |
In this research, we mainly proposed and experimentally investigated a method for automatically assigning highly accurate teacher labels on medical images without lesion region teacher labels by fusing lesion information from medical record data into a supervised machine learning framework.
The proposed method estimates lesion area teacher labels from extremely rough and weak teacher data, consisting of lesion position coordinates and lesion length diameter, by using multiple 3D U-Nets. Validation experiments by CT image data, including lung nodules, showed that the estimated teacher labels with good Dice coefficients were obtained for all three types of nodules: solid nodules, GGO nodules, and Mixed-GGO nodules.
We also investigated pixel discrimination, which is essential for lesion area extraction. A multi-species lesion detection process on FDG-PET/CT images based on unsupervised pixel anomaly detection was proposed and published as a journal article.
|
| Free Research Field |
画像診断支援
|
| Academic Significance and Societal Importance of the Research Achievements |
本研究は、画像診断支援AIシステム開発において不可欠なデータである"病変領域教師データ"の生成コストを大幅に削減することに寄与する可能性を秘めた研究である。この病変領域教師データは、専門知識を持つ放射線診断医が画素単位で手入力することでしか作成することができない。よって、病変領域教師データの作成および収集における人的・時間的コストは極めて大きい。本研究は、このようなデータ作成を少数に抑えることができ、かつ病院に蓄積された診断データを教師ラベルの作成作業無しに後ろ向き研究利用するための方策となりうる。この技術を用いることで診断支援AIシステムの開発速度向上が見込めることから、社会的意義は高い。
|
