Convolutional Neural Networks Accurately Predict Benign versus Malignant Status Among Peripheral Nerve Sheath Tumors

Alexander Mazal, Liyuan Chen, Feng Poh, Jing Wang, Michael Folkert, Oganes Ashikyan, Parham Pezeshk and Avneesh Chhabra

Abstract
Background: Peripheral nerve sheath tumors (PNSTs) comprise ~5-10% of soft tissue tumors encountered in the clinical setting. Benign lesions (BPNSTs), such as neurofibromas and schwannomas are often asymptomatic or cause neuropathy. Malignant peripheral nerve sheath tumors (MPNSTs) frequently exhibit rapid invasive behavior and metastatic spread. MR imaging markers do not reliably differentiate BPNSTs from MPNSTs. Convolutional neural networks employ machine learning and multi-order statistics to derive imaging signatures that could improve diagnostic assessment of PNSTs.
Purpose: To evaluate whether convolutional neural networks can accurately differentiate BPNSTs from MPNSTs and compare the accuracy to that of expert radiologist interpretation.
Materials and Methods: MR images from 47 patients with histologicallyconfirmed PNSTs were identified. Two separate convolutional neural networks (CNNs) were created using fat-suppressed T2-weighted (fsT2W) images alone (CNN 1), and fsT2W images in combination with pre- and post-contrast T1- weighted imaging (CNN 2). CNN performance was compared to interpretation by two experienced radiologists.
Results: CNN 1 performed comparably to the radiologists, achieving an accuracy and area under the curve (AUC) of 87% and 0.89, respectively. By comparison, radiologist 1 and 2 achieved accuracies and AUCs of 73%, 0.83 and 93%, 0.83, respectively. No significant differences were found between the accuracies or AUCs of either radiologist and CNN 1 (p>0.05). CNN 2 achieved an accuracy and AUC of 93%, 0.94. Using all image sequences together, radiologists 1 and 2 achieved accuracies and AUCs of 71%, 0.81 and 71%, 0.70, respectively.
Conclusion: Convolutional neural networks accurately differentiated BPNSTs versus MPNSTs in our investigation. Larger studies may be needed to validate these results.

Published on: February 22, 2021
doi: 10.17756/jnpn.2021-038
Citation:  Mazal A, Chen L, Poh F, Wang J, Ashikyan O, et al. 2021. Convolutional Neural Networks Accurately Predict Benign versus Malignant Status Among Peripheral Nerve Sheath Tumors. J Neuroimaging Psychiatry Neurol 6(1): 1-6.

195
Downloads