WEKO3
アイテム
Four-dimensional computed tomography artifact correction using deep neural network
https://repo.qst.go.jp/records/66913
https://repo.qst.go.jp/records/669136415a8a3-aabb-405e-95dd-336995e1d1eb
| Item type | 会議発表用資料 / Presentation(1) | |||||
|---|---|---|---|---|---|---|
| 公開日 | 2018-09-18 | |||||
| タイトル | ||||||
| タイトル | Four-dimensional computed tomography artifact correction using deep neural network | |||||
| 言語 | ||||||
| 言語 | jpn | |||||
| 資源タイプ | ||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_c94f | |||||
| 資源タイプ | conference object | |||||
| アクセス権 | ||||||
| アクセス権 | metadata only access | |||||
| アクセス権URI | http://purl.org/coar/access_right/c_14cb | |||||
| 著者 |
森, 慎一郎
× 森, 慎一郎× 平井, 隆介× 坂田, 幸辰× 森 慎一郎× 平井 隆介× 坂田 幸辰 |
|||||
| 抄録 | ||||||
| 内容記述タイプ | Abstract | |||||
| 内容記述 | Purpose Use of multi-slice CT could degrade a quality of 4DCT images due to irregular breathing. To improve the image quality, we developed three types of deep neural network (DNNs). \nMaterials and Methods A total 857 thoracoabdominal 4DCT data was acquired using 320-slice CT. Since 320-slice CT obtained 16cm scan region1 in a single rotation, there is no 4DCT artifact (= reference image). To simulate 4DCT acquired by multi-slice CT, we resorted 4DCT images at respective couch positions randomly by replacing from other respiratory phases. Due to the limitation of GPU memory, however, it is impossible to import 3D/4DCT data sets into the DNN, therefore, a single or multiple image in coronal section were imported. Image generator DNN was based on the convolutional autoencoder with skip connections. The 1st type DNN was trained to be close to the quality of the output image to that of the reference image by importing a single image. The artifacts might be corrected by selecting the image regions in other respiratory phases. Therefore, each single image at respective respiratory phases was imported into the 2nd type. Both DNNs could correct artifacts at different slice positions independently, as a result, the geometrical accuracy might be degraded in the sagittal section. The 3rd type was included the artifact discriminator and the image generator. The discriminator was based on ResNet50 and trained to detect artifact regions. \nResults All DNNs successfully corrected the artifacts on the coronal images. However, the geometrical accuracy in the sagittal sections was improved especially in the 3rd type DNN. Diaphragm position was more accurate than others by applying the 3rd DNN because the artifact map supported the correction position information. \nConclusion By providing the correction position, the DNN based 4DCT artifact correction accuracy could be improved. |
|||||
| 会議概要（会議名, 開催地, 会期, 主催者等） | ||||||
| 内容記述タイプ | Other | |||||
| 内容記述 | 第１１６回日本医学物理学会学術大会 | |||||
| 発表年月日 | ||||||
| 日付 | 2018-09-15 | |||||
| 日付タイプ | Issued | |||||
