Construction of prediction model of lung adenocarcinoma by machine learning
| Project/Area Number |
17K08740
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Research Field |
Human pathology
|
|---|
| Research Institution | Kyoto University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2017-04-01 – 2020-03-31
|
| Project Status |
Completed (Fiscal Year 2019)
|
| Budget Amount *help |
¥4,810,000 (Direct Cost: ¥3,700,000、Indirect Cost: ¥1,110,000)
Fiscal Year 2019: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
Fiscal Year 2018: ¥1,820,000 (Direct Cost: ¥1,400,000、Indirect Cost: ¥420,000)
Fiscal Year 2017: ¥2,080,000 (Direct Cost: ¥1,600,000、Indirect Cost: ¥480,000)
|
|---|
| Keywords | 肺癌 / 腺癌 / 予後予測 / 機会学習 / 深層学習 / 機械学習 / 肺腺癌 / 病理組織像 / ウェーブレット変換 / 畳み込みニューラルネットワーク / デジタル画像解析 |
|---|
| Outline of Final Research Achievements |
As the first step of the study, we performed digital pathological image analysis by multi-resolution analysis and improved kNN method. After extracting labeled patch images (128x128 pixels) from the annotated region of interest as training data, they were wavelet transformed and digitized by cluster analysis (11 variables). Next, multi-resolution analysis was performed using a support vector machine. As a result, it was possible to distinguish between tumor and non-tumor with an accuracy rate of 0.7777. To improve the result, we developed a novel deep learning method (Soft switch FCN method) and conducted a study using the system. With the proposed method, the accuracy rate for the determination between tumor and non-tumor was 0.95. On the other hand, the accuracy rate for each class was not satisfactory results. We are keeping to examine the proposed algorithm with the total dataset to construct Ca-CHS that can predict the prognosis of the patients with resected lung adenocarcinoma.
|
| Academic Significance and Societal Importance of the Research Achievements |
病理デジタル画像においても深層学習を用いた判別の可能性が示せた。しかしながら，教師データであるWSI画像は放射線画像と比較し大きく，そのまま，入力することはできない。そのため，パッチ画像として入力することになるが，組織構築の判別を行う際はより広い視野をパッチとしなければならないためその画質は落とさざるを得ないジレンマがある。そこで我々はU-Netに，最適な視野領域の組み合わせを予測できるSoft switch法を組み合わせたSoft switch FCN法を共同開発し，検討に用いた。肺癌の予後予測に関しては検討途中であるが，組織構造の分類には有用である可能性が示され，その意義は大きいと考える。
|
Report
(4 results)
Research Products
(5 results)
