キクチ　マサユキ   MASAYUKI KIKUCHI   菊池　眞之    所属   コンピュータサイエンス学部 コンピュータサイエンス学科    職種   准教授
言語種別	英語
発行・発表の年月	2023/10
形態種別	国際会議論文
査読	査読あり
標題	Automatic Detection of Spine Region using Multiple Pseudo 3D U-Net Models with Weighted Average Voting and Attention Mechanisms
執筆形態	共著
掲載誌名	ICFIP2023(International Conference on Frontiers of Image Processing)
掲載区分	国外
出版社・発行元	ICFIP
総ページ数	6
担当区分	最終著者
著者・共著者	Kai Yang, Kanon Kobayashi, Masayuki Kikuchi
概要	In the field of medical imaging, the U-Net architecture has emerged as a prominent deep learning model owing to its proficiency in handling segmentation tasks. A previous study by Shigeta et al. exploited the U-Net for pseudo 3D feature extraction. In our present study, we used supervised learning aiming to automate CT image extraction. To this end, we deployed convolutional neural networks (CNNs), particularly specialized for medical image processing. Our approach adopted Shigeta's U-Net model, implementing a model that applies average calculations to data within a n x n x n range located three units distant from the coordinates, with the average output derived from the weights obtained from both range of 1×1×1 and 7×7×7 models. Furthermore, each of these models are added with attention gates, facilitating pseudo 3D feature extraction by assigning higher weights to the most pertinent features, thereby potentially enhancing model performance, and by utilizing ensemble learning, we succeeded in augmenting model diversity and mitigating the issue of overfitting.