Comparison of a STIR- and T1-based radiomics model to differentiate between plexiform neurofibromas and malignant peripheral nerve sheath tumors in neurofibromatosis type 1 (NF1)
Ina Ly1, Tianyu Liu1, Wenli Cai1, Olivia Michaels1, Maya Viera1, Daniel Kwon1, Miriam Bredella1, Justin Jordan1, Dana Borcherding2, Demetrius Boswell3, Crystal Burgess 4, Ping Chi5, Peter de Blank3, Eva Dombi4, Angela Hirbe2, Bruce Korf6, Shernine Lee6, Victor Mautner7, Mairim Melecio-Vazquez5, Zachary Mulder8, Kai Pollard9, Christine Pratilas9, Johannes Salamon7, Divya Srihari2, Matthew Steensma8, Brigitte Widemann4, Jaishri Blakeley10, Scott Plotkin1
1Massachusetts General Hospital, 2Washington University, 3Cincinnati Children's Hospital Medical Center, 4National Cancer Institute, 5Memorial Sloan Kettering Cancer Center, 6University of Alabama At Birmingham, 7University of Hamburg, 8Van Andel Research Institute, 9Johns Hopkins University, 10Johns Hopkins University School of Medicine
Objective:

Evaluate the performance of radiomics to classify plexiform neurofibromas (PNF) and malignant peripheral nerve sheath tumors (MPNST) in NF1

 

Background:

MPNST are the leading cause of death in NF1 and most often arise from a pre-existing benign PNF. Short TI inversion recovery (STIR) sequences on MRI display tumor extent but do not permit accurate differentiation between PNF and MPNST. Furthermore, STIR sequences are not routinely acquired for certain body regions. T1-weighted pre-contrast (T1) sequences are more standardly obtained. We developed a radiomics model using STIR and T1 sequences to differentiate between NF1-associated PNF and MPNST.

Design/Methods:

136 PNF and 91 MPNST from nine centers were segmented on STIR sequences (if available) or T2-weighted fat-saturated or T1-weighted fat-saturated post-contrast sequences. Segmentations were co-registered to T1 sequences. Standard pre-processing included N4 bias field correction, intensity normalization (mean 120 SI, SD 80 SI), and resampling (1 mm3 voxel resolution). 107 radiomic features were extracted from STIR- and T1-derived segmentations using PyRadiomics. We applied the Boruta algorithm and correlation removal for selection of important features. A Random Forest model was built using the top ten selected features. The data were divided into a training/validation and test set (7:3 ratio). Ten-fold cross-validation was performed and repeated 100 times. Model performance was evaluated using the AUC.

Results:

The AUC (95% confidence intervals) for the test set was 0.807 (0.694-0.921), 0.744 (0.614-0.875), and 0.836 (0.705-0.968) for the STIR, T1, and combined STIR+T1 model, respectively. For the STIR+T1 model, nine texture features and one shape feature were included in the model.

Conclusions:

Using clinical MRIs, our radiomics models demonstrate high performance in classifying PNF and MPNST on STIR and STIR+T1 sequences. Our inclusion of multicenter MRIs enhances model generalizability. These models can potentially be integrated into the clinical workflow to help clinicians in the early identification of MPNST or pre-malignant atypical neurofibromas.

 

10.1212/WNL.0000000000203702