COT-16 Development of automatic lesion extraction application using artificial intelligence for the purpose of simplifying tumor volume measurement of meningioma. (6th December 2021)

• Record Type: Journal Article
• Authors: Hirayama, Ryuichi; Iwata, Takamitsu; Yamada, Shuhei; Kuroda, Hideki; Nakagawa, Tomoyoshi; Kijima, Noriyuki; Okita, Yoshiko; Kagawa, Naoki; Kishima, Haruhiko
• Is Part Of: Neuro-oncology advances. Volume 3(2021)Supplement 6
• Page Start: vi30
• Page End: vi30
• Publication Date: 2021-12-06
• Subjects: Meningioma -- Automated volumetry -- Artificial intelligence; 616.99481
• DOI: 10.1093/noajnl/vdab159.118
• Languages: English
• ISSNs: 2632-2498

Abstract

Abstract: BACKGROUND: With the widespread use of MRI equipment and brain scans, opportunities to perform follow-up examinations for meningiomas have increased. On the other hand, an objective evaluation index for meningiomas characterized by slow changes on imaging has not been established. To establish a volume-based evaluation index for meningoceles, we are developing an application for automatic lesion extraction using artificial intelligence as a highly reproducible tumor volume measurement technique that enables large volume image data processing. METHODS: In this study, 195 patients with meningioma who underwent contrast-enhanced MRI imaging at Osaka University Hospital were included. The images were manually extracted by three neurosurgeons and used as supervised data. deeplabV3 was used as the learning network. All the supervised data were randomly divided into training (80%) and testing (20%) data, and the application was constructed by deep learning and validation with 5-fold cross-validation. The matching rate of the area of the region automatically extracted by the device against the test data and the mean square error rate of the calculated tumor volume were used as indices of the product measurement performance. RESULTS: The matching rate using the automatic extraction application for the correct data(Dice index) was 91.5% on average. The mean squared error rate of the tumor volume calculated from these extracted regions was 8.84%. CONCLUSION: We consider that this application using artificial intelligence has a certain degree of validity in terms of the accuracy of extracted lesions. In the future, it is necessary not only to improve the performance of the equipment but also to clarify the clinical significance of the new imaging biomarkers based on tumor volume that can be obtained from these lesion extraction techniques.