{"created":"2023-05-15T14:59:38.687385+00:00","id":80928,"links":{},"metadata":{"_buckets":{"deposit":"4101ef2b-c75c-4cd5-8b52-318b6bba4978"},"_deposit":{"created_by":1,"id":"80928","owners":[1],"pid":{"revision_id":0,"type":"depid","value":"80928"},"status":"published"},"_oai":{"id":"oai:repo.qst.go.jp:00080928","sets":["1"]},"author_link":["1003022","1003023","1003019","1003024","1003020","1003021"],"item_8_biblio_info_7":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicIssueDates":{"bibliographicIssueDate":"2020-11","bibliographicIssueDateType":"Issued"},"bibliographicPageEnd":"158","bibliographicPageStart":"151","bibliographicVolumeNumber":"80","bibliographic_titles":[{"bibliographic_title":"Physica Medica"}]}]},"item_8_description_5":{"attribute_name":"抄録","attribute_value_mlt":[{"subitem_description":"機械学習を用いて3次元CT画像から４次元CT画像を生成する技術について記載している。\nPurpose/Objective(s)\nOur markerless tumor tracking algorithm required 4DCT data to train model file. To perform the markerless tracking for respiratory gated treatment even 4DCT is not available, we developed deep neural network (DNN) to predict 4DCT from 3DCT.\nMaterials/Methods\nA total of 2420 thoracic 4DCT data sets (436 patients) was used to train DNN, which was designed to export 9 different deformed vector fields (DVFs) from a single CT data based on the 3D convolutional autoencoder with shortcut connections. However, since our DNN input a single data (CT image) and output 9 different data (DVF), it is not good to use shortcut connection in between encoder block and decoder block like U-net. To solve this problem, shortcut connection was applied to every two convolutional layers (residual block). Since DNN was trained to export 9 DVF files close to those of the ground-truth DVFs, we used the multitask learning technique. DVF from exhale to respective phases was calculated by applying deformable image registration. Then 3DCT data at exhale was transformed by using the predicted DVFs to obtain simulated 4DCT data. We compared markerless tracking accuracy in between original and simulated 4DCT and evaluated for 20 lung patients. Our tracking algorithm uses machine learning approaches. For the training stage, a pair of digitally reconstructed radiography images was calculated by using 4DCT. Training data was divided into subimages around the tumor region (including tumor and not). Subimage feature information was expressed by the intensity gradient and tracking algorithm parameters were optimized by the randomized tree method to calculate tumor likelihood map. Use of tumor positions as a function of respiratory phase designed regression model to calculate a tumor probability estimation map. For the predicting stage, incoming fluoroscopic image was cropped to a subimage, and the tracking algorithm calculated a tumor likelihood map and a tumor probability estimation map. This process was repeated by sliding subimage position on the incoming fluoroscopic images. Then the final tumor position on fluoroscopic image was then estimated by multiplying likelihood map and probability map.\nResults\nSimulated 4DCT for respective phases were successfully generated from 4DCT at T50 and seemed natural close to the original 4DCT. Tracking positional errors averaged over all patients were -.0.07±0.25 mm, -0.04±0.34 mm, and -0.36±0.46 mm for left-right (LR), anterior-posterior (AP), and superior-inferior (SI) direction, respectively. The 95% percentile averaged over all patient was 0.26 mm, 0.36 mm, and 0.22 mm in RL, AP and SI directions, respectively.\nConclusion\nOur developed DNN for generating simulated 4DCT data is useful for the markerless tumor tracking when original 4DCT is not available. Use of our DNN would accelerate markerless tumor tracking in radiotherapy and increased treatment accuracy in thoracoabdominal treatment.","subitem_description_type":"Abstract"}]},"item_8_publisher_8":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"Elsevier"}]},"item_8_relation_14":{"attribute_name":"DOI","attribute_value_mlt":[{"subitem_relation_type_id":{"subitem_relation_type_id_text":"10.1016/j.ejmp.2020.10.023","subitem_relation_type_select":"DOI"}}]},"item_8_relation_17":{"attribute_name":"関連サイト","attribute_value_mlt":[{"subitem_relation_type_id":{"subitem_relation_type_id_text":"https://www.sciencedirect.com/science/article/pii/S1120179720302660","subitem_relation_type_select":"URI"}}]},"item_8_source_id_9":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"1120-1797","subitem_source_identifier_type":"ISSN"}]},"item_access_right":{"attribute_name":"アクセス権","attribute_value_mlt":[{"subitem_access_right":"metadata only access","subitem_access_right_uri":"http://purl.org/coar/access_right/c_14cb"}]},"item_creator":{"attribute_name":"著者","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Mori, Shinichiro"}],"nameIdentifiers":[{"nameIdentifier":"1003019","nameIdentifierScheme":"WEKO"}]},{"creatorNames":[{"creatorName":"Hirai, Ryusuke"}],"nameIdentifiers":[{"nameIdentifier":"1003020","nameIdentifierScheme":"WEKO"}]},{"creatorNames":[{"creatorName":"Sakata, Yukinobu"}],"nameIdentifiers":[{"nameIdentifier":"1003021","nameIdentifierScheme":"WEKO"}]},{"creatorNames":[{"creatorName":"Shinichiro, Mori","creatorNameLang":"en"}],"nameIdentifiers":[{"nameIdentifier":"1003022","nameIdentifierScheme":"WEKO"}]},{"creatorNames":[{"creatorName":"Ryusuke, Hirai","creatorNameLang":"en"}],"nameIdentifiers":[{"nameIdentifier":"1003023","nameIdentifierScheme":"WEKO"}]},{"creatorNames":[{"creatorName":"Yukinobu, Sakata","creatorNameLang":"en"}],"nameIdentifiers":[{"nameIdentifier":"1003024","nameIdentifierScheme":"WEKO"}]}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"eng"}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourcetype":"journal article","resourceuri":"http://purl.org/coar/resource_type/c_6501"}]},"item_title":"Simulated four-dimensional CT for markerless tumor tracking using a deep learning network with multi-task learning","item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"Simulated four-dimensional CT for markerless tumor tracking using a deep learning network with multi-task learning"}]},"item_type_id":"8","owner":"1","path":["1"],"pubdate":{"attribute_name":"公開日","attribute_value":"2020-10-28"},"publish_date":"2020-10-28","publish_status":"0","recid":"80928","relation_version_is_last":true,"title":["Simulated four-dimensional CT for markerless tumor tracking using a deep learning network with multi-task learning"],"weko_creator_id":"1","weko_shared_id":-1},"updated":"2023-05-15T19:31:52.745155+00:00"}