NI-07 VALIDATION OF MACHINE LEARNING BASED HIGH GRADE GLIOMA MR SEGMENTATION VIA METHIONINE PET. (16th December 2019)

• Record Type: Journal Article
• Authors: Kinoshita, Manabu; Ozaki, Tomohiko; Arita, Hideyuki; Kagawa, Naoki; Kanemura, Yonehiro; Fujimoto, Yasunori; Sakai, Mio; Watanabe, Yoshiyuki; Nakanishi, Katsuyuki; Shimosegawa, Eku; Hatazawa, Jun; Kishima, Haruhiko
• Is Part Of: Neuro-oncology advances. Volume 1(2019)Supplement 2
• Page Start: ii27
• Page End: ii27
• Publication Date: 2019-12-16
• Subjects: 616.99481
• DOI: 10.1093/noajnl/vdz039.120
• Languages: English
• ISSNs: 2632-2498

Abstract

Abstract: Treatment planning and lesion-follow up are generally conducted by contrast-enhanced MRI in glioma patient care. On the other hand, there are, however, substantial concerns whether MRI actually reflects the extension or activity of this neoplasm, which information should be fundamentally important at every step when treating this disease. As a matter of fact, the authors of this investigation have already shown that there is no difference in tumor cell density within areas with and without contrast enhancement (J Neurosurg. 2016, 125(5):1136–1142.) and furthermore that the geometry of MRI based-radiation treatment planning is significantly altered when methionine PET is integrated for this purpose (J Neurosurg. 2018 published on-line). Regardless of these concerns, there is great interest in the research community to construct a machine learning based fully automated brain tumor segmentation tool specific for gliomas using MRI. The authors attempted to validate this method by comparing MRI-based automated brain tumor segmentation and methionine PET. Consecutively collected 45 high-grade gliomas (GBM-26, grade3-19) were analyzed. BraTumIA, an automated brain tumor segmentation tool, was used for machine learning based lesion segmentation. At the same time, lesions were segmented using various thresholds on methionine PET. The authors observed 40% of pseudo-positive and 90% of pseudo-negative error on BraTumIA based lesion segmentation when methionine PET was considered as ground truth with a cut-off of 1.3 in T/N ratio. Pseudo-negative error was as high as 60% even if the threshold was elevated to 2.0. Although machine learning based glioma segmentation is expected to expand in both research and clinical use, the observed results caution the use of MRI as ground truth of spatial extension of glioma and researchers should be reminded that this imaging modality may obscure the true behavior of the disease within the patient in some cases.