Neurofibromatosis type 1 (NF1) is a genetic multi-system disease that predisposes affected individuals to the formation of neurofibromas, a group of benign tumors that originate from the peripheral nerve sheath. Internal neurofibromas (iNFs) are located deep within the body and affect up to 60% of individuals with NF1. Growth of iNFs over time can cause significant complications, including pain, neurologic dysfunction, and compression of other important organs. Rapidly growing iNFs may also herald malignant transformation of a benign tumor into a malignant peripheral nerve sheath tumor (MPNST) – an aggressive type of sarcoma with limited treatment options. For these reasons, it is important for clinicians to be able to identify iNFs that carry an increased risk of future growth to prevent these complications from occurring in the first place. Since iNFs are located deep within the body and may be asymptomatic, the only way to know about their presence, number, and size is by imaging. The most accurate method to visualize iNFs is whole-body magnetic resonance imaging (WBMRI), which captures total-body tumor burden.

Preliminary data suggest that certain manually designed, descriptive MRI features (such as tumor size greater than or equal to 3 cm, tumor border characteristics, and intratumoral MRI signal characteristics) are associated with more rapid tumor growth. However, the generalizability and applicability of these MRI features to patients of different age groups with variable clinical characteristics is currently unknown. Descriptive MRI features are also based on subjective assessments and may be difficult to reproduce. Last, descriptive features may ignore complex but important imaging features that are not visible to the human eye. In contrast to descriptive features, deep learning – a form of machine learning – is based on computer algorithms that learn directly from MR images in an unbiased way. Deep learning may thus identify complex relationships among imaging patterns, clinical data, and tumor growth behavior that are not captured by manually designed, descriptive features.

The primary objective of this proposal is to develop and test two distinct and clinically deployable tumor growth prediction models that identify iNFs at risk of future growth, based on either manually designed, descriptive MRI features (Model 1) or deep learning (Model 2). To accomplish this, we will first collect WBMRIs and clinical data from approximately 185 adults and children with NF1 and an estimated 1095 iNFs from three major NF1 referral centers. Approximately 80% of tumors will be used to develop the models while the remaining 20% will be used to validate and confirm the models’ ability to predict tumor growth. Clinical and genetic data will also be included in the development of both models. Based on the unbiased nature of Model 2 and the ability of deep learning to detect complex relationships between data, our hypothesis is that Model 2 will demonstrate better performance in predicting future iNF growth than Model 1.

We anticipate that, by the end of this 3-year award period, the availability of an iNF growth prediction model will permit identification of patients at high risk of future tumor growth, which will enable enhanced surveillance and patient selection for treatment. Conversely, this research will also help identify patients at low risk of tumor growth who may not require surgery or chemotherapy, thus sparing them costly and unnecessary treatment interventions. In the long term, this will prevent both tumor- and treatment-related morbidity and mortality. Given the inclusion of both adults and children from multiple centers in our model development, we anticipate that our models can be applied to a diverse group of patients across different institutions. If our hypothesis is true, our research will lead to development of a predictive model that can integrate complex imaging and clinical and genetic data, which is relevant in a heterogeneous disease like NF1. If Model 1 proves to be better than Model 2, it will provide a cost-effective and practical method for tumor growth prediction that can be easily implemented across centers. Other long-term applications derived from this research include the extension of the deep learning-based model to build a model that can predict future malignant transformation or the likelihood of response to chemotherapy in NF1. In addition, the model may also be applied to study tumor growth behavior in related neurogenetic tumor predisposition syndromes such as NF2 and schwannomatosis.