Towards Improved Medical Imaging Analysis for Oncology Care: Clinical Experience with Simulation Platform for Brain Tumors

Abstract

Analyses of large scale medical imaging data involving machine learning are rapidly evolving to include diagnosis, treatment and prognosis in oncology care. In neuro-oncology, radiomics and similar multi-parameter analysis approaches are advancing clinical trajectories. A common challenge to robust machine learning approaches is the lack of “ground truth” datasets for supervised learning upon which to base robust prediction models. The ability to simulate the mechanistic aspects (geometry, location, physical properties etc.) of lesions/tumours may significantly boost the ability to assess, benchmark and compare competing analysis techniques for identification of the most relevant techniques for clinical care improvements. In this study, we propose and evaluate a simulation toolset for enabling the creation of large numbers of “ground truth” datasets and make the implementation available for other groups/researchers/clinicians to use. The simulation tool is developed using the Feature Manipulation Engine (Safe Software, Surrey, Canada), which enables the integration, validation, transformation and processing of over 350+ data formats. Based on clinical knowledge and experience, canonical brain tumours (both high- and low-grade gliomas) were defined by specifying location, signal intensity, shape and other characteristics. Magnetic resonance imaging (MRI) datasets with simulated brain tumours were generated by inserting tumours with prescribed properties into the template structural brain in MNI space. Blinded clinical and medical imaging experts assessed the quality and visual accuracy for outcome measures. The simulation toolset successfully incorporated the simulated tumours within specified locations in the brain MRI using an easy-to-use graphical user interface. Initial results suggest the verification of the visual realism of the simulated datasets by clinical and medical imaging experts. On-going analyses are further assessing the validity of the tool with quantitative analysis . The new simulation toolset is capable of creating simulated MR images incorporating mechanistic aspects of common brain tumours. This may enable the creation of large numbers of “ground truth” datasets for assessing medical imaging analysis techniques. Furthermore, the easy availability of the tool to everyone may spur the wide-scale use of true “ground truth” data for further development of novel techniques and assessing imaging based biomarkers for improved clinical care.