(P 154) PREOPERATIVE CLASSIFICATION OF NERVE SHEATH TUMORS USING RADIOMICS-BASED MACHINE LEARNING

Objective: Malignant peripheral nerve sheath tumors (MPNSTs) are aggressive soft tissue sarcomas occurring in neurofibromatosis type 1 (NF1) patients. High-grade MPNSTs carry a high risk of metastasis and recurrence, making early recognition and surgical resection crucial for improved survival. The resection of high-grade MPNSTs often leads to postoperative complications, while neurofibromas can be adequately resected with minimal nerve damage. Therefore, preoperative differentiation between benign peripheral nerve sheath tumors (BPNSTs) and MPNSTs is important. However, current imaging tools may not always provide sufficient diagnostic accuracy, leading to the need for biopsies and associated burdens. Clear distinguishing features in magnetic resonance imaging (MRI) are not yet known. Radiomics provides a potential new tool in the diagnostic armamentarium, which may reduce the need for biopsies. This study aims to develop a radiomics model that utilizes quantitative imaging features and machine learning to differentiate BPNSTs from MPNSTs based on T1- and T2-weighted MRI sequences.

Methods: Clinical features and diagnostic MRIs were collected from BPNST and MPNST patients at our tertiary sarcoma referral center between 2000-2019. Lesions were manually and semi-automatically segmented on the MRI scans where the tumor was best visible. Segmentations were warped to the other sequences using image registration. For each lesion, on each sequence, 564 radiomics features were extracted. For classification, the Workflow for Optimal Radiomics Classification (WORC) algorithm was used, which includes a large set of commonly used radiomics methods and uses automated machine learning to determine their optimal combination based on the training set. Evaluation was performed using a 100x random-split cross-validation with, in each iteration, 20% of the data for testing.

Results: A total of 35 MPNSTs and 74 BPNSTs were included. The T1-weighted (T1w) MRI radiomics model outperformed all the other radiomics models with imaging only, with an area under the curve (AUC) of 0.71. The incorporation of additional MRI scans to the T1w MRI model did not enhance the performance. The radiomics model combining T1w MRI and clinical features achieved an AUC of 0.74. Two experienced radiologists visually scoring the tumors achieved AUCs of 0.75 and 0.66, respectively.

Conclusion: Radiomics based on T1w MRI scans and clinical features shows some degree of ability to distinguish MPNSTs from BPNSTs, which holds potential to aid in the management of these tumors.