NIMG-16. COMPARISON OF A STIR- AND T1-WEIGHTED-BASED RADIOMICS MODEL TO DIFFERENTIATE BETWEEN PLEXIFORM NEUROFIBROMAS AND MALIGNANT PERIPHERAL NERVE SHEATH TUMORS IN NEUROFIBROMATOSIS TYPE 1 (NF1)

BACKGROUND

Plexiform neurofibromas (PNF) and malignant peripheral nerve sheath tumors (MPNST) are best visualized on short TI inversion recovery (STIR) sequences on MRI. However, STIR sequences are not routinely acquired in the clinical setting. T1-weighted pre-contrast (T1W) sequences are more standardly obtained but provide insufficient contrast for tumor identification. We developed a radiomics model based on STIR and T1W sequences to differentiate between NF1-associated PNF and MPNST.

METHODS

Using a 3D quantitative imaging analysis software (3DQI), 68 MPNST and 79 PNF from 134 participants at nine centers were segmented on STIR sequences (if available) or T2 fat-saturated or T1-weighted fat-saturated post-contrast sequences. Tumor regions of interest were co-registered to T1W 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 using PyRadiomics. To classify tumors as PNF or MPNST, we applied the Boruta algorithm and correlation removal for selection of important features. A Random Forest model was built using the top five selected features. The data were divided into a training/validation and test set (7:3 ratio). Five-fold cross-validation was performed and repeated 100 times. Model performance was evaluated using AUC, sensitivity, specificity, accuracy, and 95% CI.

RESULTS

For the STIR-based model, AUC was 0.856 (95% CI 0.727-0.984), sensitivity 0.6, specificity 0.833, and accuracy 0.727 in the test set. For the T1W-based model, AUC was 0.867 (95% CI 0.743-0.990), sensitivity 0.8, specificity 0.79, and accuracy 0.794 in the test set.

CONCLUSIONS

Our radiomics models demonstrate high and comparable performance to distinguish between PNF and MPNST on STIR and T1W sequences. Our inclusion of multicenter MRIs enhances model generalizability. These models can potentially be integrated into the radiologic workflow to help clinicians in the early identification of MPNST or pre-malignant atypical neurofibromas on clinical MRIs.

NIMG-08. A MULTI-CENTER RADIOMICS-BASED MODEL TO DIFFERENTIATE BETWEEN NEUROFIBROMATOSIS TYPE 1-ASSOCIATED PLEXIFORM NEUROFIBROMAS AND MALIGNANT PERIPHERAL NERVE SHEATH TUMORS

BACKGROUND

Several MRI features are proposed to distinguish between plexiform neurofibromas (PNF) and malignant peripheral nerve sheath tumors (MPNST) in neurofibromatosis type 1 (NF1), including tumor size, margins, and degree of heterogeneity. However, most of these features are descriptive in nature, subject to intra-/interrater variability, and based on small single-institution studies. The goal of this study was to identify radiomic features that can differentiate between NF1-associated PNF and MPNST.

METHODS

31 MPNSTs and 24 PNFs from five centers were segmented on short TI inversion recovery sequences using a semi-automated segmentation software (3DQI). Standard pre-processing was performed, including N4 bias field correction, intensity normalization (using a mean of 120 SI and standard deviation of 80 SI), and resampling to 1 mm3 voxel resolution. 1688 radiomic features were extracted from the tumor region of interest using PyRadiomics, an open-source Python radiomics package. To classify tumors as PNF or MPNST, we implemented the Boruta algorithm and correlation removal for selection of important features. A Random Forest model was built using the top ten selected features. Five-fold cross-validation was performed and repeated 100 times. Model performance was evaluated using the area under the ROC curve (AUC), sensitivity, specificity, accuracy, and confidence intervals.

RESULTS

The top ten features included in the model were five intensity features, two shape features, and three texture features. The model demonstrated an AUC of 0.891 (95% CI 0.882-0.899), sensitivity of 0.744, specificity of 0.847, and accuracy of 0.802 (95% CI 0.792-0.813).

CONCLUSIONS

Our machine learning model demonstrated high performance in classifying tumors as either PNF or MPNST in NF1 individuals. Inclusion of additional tumors for model training and testing on an independent dataset are underway. Ultimately, our model may enable improved differentiation between PNF and MPNST compared to descriptive MRI features, permit early patient risk stratification, and improve patient outcomes.

INNV-04. A MULTI-INSTITUTIONAL CLINICAL AND MRI REPOSITORY OF NEUROFIBROMATOSIS TYPE 1-ASSOCIATED PERIPHERAL NERVE SHEATH TUMORS

BACKGROUND

Individuals with neurofibromatosis type 1 (NF1) frequently have peripheral nerve sheath tumors (PNST), including plexiform neurofibromas (PNF), atypical neurofibromas (ANF), and malignant peripheral nerve sheath tumors (MPNST). These tumors reflect a histologic spectrum from benign to malignant. Various clinical and MRI-based features are proposed as risk factors for MPNST development based on small single-institution studies. A major barrier to study these risk factors is collation and annotation of multi-center serial MRIs. To address this, we created a standardized database of clinical data and longitudinal MRIs from NF1-associated PNST from nine international NF1 referral centers.

METHODS

Clinical data from NF1 patients are collected in Research Electronic Data Capture databases housed at Massachusetts General Hospital and Washington University, including demographic information, genotype, disease course, treatment history, and survival. ANF and MPNST require histologic confirmation whereas a diagnosis of PNF can also be made based on clinical/radiographic stability. Longitudinal MRIs predating the histologic diagnosis are uploaded to a HIPAA-compliant cloud-based system.

RESULTS

Data from 200 patients (87 females, 113 males) with 217 tumors (75 PNF, 40 ANF, 102 MPNST) have been collected. 280 regional and 108 whole-body MRIs have been identified. Median age at the time of histologic diagnosis is 30 years (range 5-64). All tumors are histologically confirmed except for 6 PNF which remained stable over time. Median follow-up time is 32 months. Of 147 patients with available survival data, 32 (21.7%) have died from MPNST progression; estimated median overall survival is 20 months.

CONCLUSIONS

In this ongoing work, we have assembled one of the largest systematically annotated clinical and MRI repositories of NF1-associated PNST from pediatric and adult NF1 patients. The data will be accessible to outside researchers which will promote interdisciplinary and multi-center collaborations. Active efforts include the identification of radiomic MRI features to differentiate between PNF and MPNST.