3次元畳み込みニューラルネットワークを用いた骨盤CT画像からの自動骨折検出

• Published in: バイオメディカル・ファジィ・システム学会大会講演論文集 Vol. 33; pp. 36 - 42
• Main Authors: Rashedur, Rahman, 丸尾, 明宏, 小橋, 昌司, 山本, 侃利, 八木, 直美, 林, 圭吾, 村津, 裕嗣
• Format: Journal Article
• Language: Japanese
• Published: バイオメディカル・ファジィ・システム学会 31.10.2020; Biomedical Fuzzy Systems Association
• Subjects: automated fracture detection; CT images; deep learning
• ISSN: 1345-1510; 2424-2586
• DOI: 10.24466/pacbfsa.33.0_36

In emergency hospitals, the automatic fracture detection system is essential for doctors and patients to recover not only the injury but also the health status without their long hospitalization. Previous studies for the fracture detection system with CT images or deep-learning have difficulty in analyzing the internal structure or confirming predicted results easily. This study proposes a system for automatic detection of pelvic fractures from 3D CT images. Firstly, it defines the labeling work as a new 3D annotation method of fractures(called 3D surface annotation). 3D shape data of pelvic bone surfaces makes the burden of it light. The feature vector inside the pelvic surface is created from 3D shape data and CT images, and learned by 3D convolutional neural networks (CNN). The proposed method was validated by using 103 subjects. Eventually, the accuracy, precision, recall and specificity for the test data were 69.5%, 60.0%, 60.4% and 75.0%