68 Views

Diffusion kurtosis and diffusion tensor based neural network model for prediction of meningioma grade: Preliminary validation results with 15 patients

Friday, May 3, 2024

Presenting Author(s)

Ivo Peto, MD

Neurosurgery Resident
Department of Neurosurgery and Brain Repair, University of South Florida
Tampa, FL, US

Co-Author(s)

YB

Yarema Bezchlibnyk, n/a

Neurosurgeon
Department of Neurosurgery and Brain Repair, University of South Florida Morsani College of Medicine, Tampa, Florida, USA
Tampa, Florida, United States
Elliot Neal, MD

Resident
University of South Florida
Tampa, FL, US
AP

Anthony Patrizz, MD

Resident
University of South Florida
Tampa, Florida, United States
EP

Elliot Pressman, MD

Resident
University of South Florida
Tampa, Florida, United States
GR

Gautam Rao, MD, N/A

Fellow
Swedish Neuroscience Institute
Tampa, Florida, United States
CR

Casey Ryan, MD

Resident
University of South Florida
Tampa, Florida, United States
KV

Kunal Vakharia, MD

Assistant Professor
University of South Florida
Tampa, Florida, United States
Hv

Harry van Loveren, MD

Professor
University of South Florida
Tampa, Florida, United States

Introduction: Radiographic classification of meningioma grade might allow for earlie, more radical surgery in higher grade tumors or early detection of progression to a higher grade. Currently, no reliable and easily accessible non-invasive means exist to achieve this goal. Aim of this study was to develop a diffusion MRI based prediction model.

Methods: A multi-shell diffusion weighted MRI acquisitions along with volumetric postcontrast T1 sequence were obtained on a 3 Tesla MRI scanner and coregistered. Diffusion tensor (DTI) and diffusion kurtosis (DKI) parametric maps were calculated using in-house made software. Tumors were manually segmented on T1 post-contrast volumetric image, which was subsequently used as a mask to obtain DKI and DTI parameters of the tumor.
Training of neural network model based on individual tumor voxels was performed utilizing DKI and DTI parametric maps obtained from 16 tumors of 16 patients (8 WHO Grade I, 8 WHO II/III), totaling 552,465 voxels. Analysis of individual tumor voxels was performed by the trained model.

Results: Internal validation on 25% randomly selected tumor voxels not included in the training yielded accuracy of 89.19% and receiver operating characteristics area under curve of 96.5% predicting tumor grade in the individual voxel. External validation on 15 meningiomas from 15 patients (8 WHO I, 7 WHO II unknown to the model (not included in the training or internal validation) was performed. Each tumor voxel was independently analyzed and the majority class voxel (WHO I vs.WHO II) within the tumor determined the overall tumor grade, yielding 93.3% accuracy. Only one misclassification (WHO I misclassified as WHO II) was noted.

Conclusion : Neural network based on combined higher order diffusion (DTI/DKI) seems to be a promising non-invasive tool for meningioma grading. Future studies are ongoing and will include a larger out-of-sample data set for training and validation.