MR Radiomic Features Improve Seizure Prediction in Children with Supratentorial Low-grade Glial Cell Tumors
Abstract number :
3.456
Submission category :
9. Surgery / 9B. Pediatrics
Year :
2024
Submission ID :
317
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Syu-Jyun Peng, PhD – Taipei Medical University
Min-Lan Tsai, MD, MS – Division of Pediatric Neurology, Department of Pediatrics, Taipei Medical University Hospital, Taipei Medical University, Taipei, Taiwan
Kevin Li-Chun Hsieh, MD – Taipei Medical University Hospital
Yen-Lin Liu, MD – Taipei Medical University Hospital
Yi-Shan Yang, MD – Taipei Medical University Hospital
Tai-Tong Wong, MD – Taipei Medical University Hospital
Rationale: Supratentorial tumors were more commonly associated with seizures than infratentorial tumors. Understanding the impact of seizures on pediatric brain tumors yield a precise diagnosis and treatment. This study aimed to investigate radiomics features between seizures and non-seizures in supratentorial low grade glial tumors in children and evaluate the value of future prediction of seizures.
Methods: Our retrospective study included a primary cohort of 48 children with supratentorial low grade (WHO grades I-II) glial cell tumor after exclusion of high-grade tumor and infratentorial tumors. All of patients received surgery or biopsy with complete MRI survey. The tumor location features and 3-D imaging features were determined between seizure (N=23) and non-seizure (N=25) groups before surgery or treatment. Two hundred and eight radiomic features were extracted from T2 FLAIR images. Tumor locations were divided into 9 parts for analyzing the significance between seizure and non-seizure group. The relationship between clinical and radiologic characteristics was also analyzed. Leave-one-out cross validation was used for prediction and radiomics and location signatures between seizure and non-seizure the value of radiomics and location separately and combination.
Results: Nine features of radiomics including 2 shapes and 7 textures were significantly different between seizure and non-seizure groups (p< 0.01). Temporal and limbic lobe involvement were significantly different between two groups (p=0.0007 and 0.013, respectively). Nine selective radiomic signatures were significantly for discriminating patients with or without seizure. The best prediction performance is 0.812 of accuracy, specificity of 0.870 and precision of 0.769 by Weighted KNN model using 9 signatures of radiomics. The accuracy and specificity reach to 0.938 and 0.88, respectively by adding the location to the radiomic model (Binary GLM logistic regression model).
Conclusions: Our results suggested that 9 features of radiomics including 2 shapes and 7 textures were significantly different between seizure and non-seizure groups. Together with significant locations, the prediction models enable more precise differentiate whether seizure or not associated with supratentorial low grade glioma.
Funding: This work was financially supported by the National Science and Technology Council, Taiwan, under the project NSTC 112-2628-E-038-001-MY3.
Surgery