Marius Staring

LUMC

In short, Marius Staring's research entails the following: 

Machine Learning to support Clinical Decision Making in Vestibular Schwannoma (MLSCHWAN)

Vestibular schwannomas (VS) are rare intracranial tumors that are (typically) benign, but may cause invalidating symptoms such as hearing loss, balance disturbance or even intracranial hypertension and brainstem compression in advanced cases. Typically, patients are monitored closely with periodic MR imaging, and only in case of tumor progression, treatment is opted for. Computational tools based on Machine Learning may allow prediction of tumor progression in an early stage, when less invasive treatment strategies are still an option. In the MLSCHWAN project our team of engineers, radiologists and ear-nose-throat surgeons, will develop tumor growth prediction tools, implement these in the clinic, and evaluate their added value.

 

 

 

MRI scan of a Vestibular Schwannoma: is it stable or does it grow?

 

Publications

• Vestibular Schwannoma Growth Prediction from Longitudinal MRI by Time-Conditioned Neural Fields
  • International Conference on Medical Image Computing and Computer-Assisted Intervention, October 2024
  • Click here to view the publication

 

Short summary of progress/results 2025

A dataset of 2,348 vestibular schwannoma patients with 10,424 MRI sessions was collected, cleaned, and uploaded to XNAT, with standardized mappings for T1 contrast-enhanced and high-quality T2 scans. A subset of 871 patients meeting inclusion criteria was selected for further analysis. Tumor segmentations were generated to assess volume, diameters, and growth rates, with manual review of those scheduled for the coming period. A baseline tumor growth prediction model, trained on 129 patients, achieved an accuracy of 0.80 but struggled with misclassification of slow-growing tumors. The DeepGrowth model, integrating neural fields and recurrent neural networks, improved tumor growth prediction, especially for complex growth patterns, by incorporating temporal encoding. Future work will focus on refining predictive models, integrating tumor texture features, and assessing clinical applicability.

 

Short summary of progress/results 2024

In this first period we have created and curated a patient inclusion list, and downloaded and pseudoanonimyzed longitudinal MR imaging data of over 2000 patients. This will be further cleaned and uploaded to an XNAT facility. We have started development of a machine learning approach based on MR imaging that predicts a binary classification of growth in the coming year of over 2 mm in diameter or less, based on a generalized radiomics approach implemented in WORC.