A clinical tool for automated detection and delineation of intracranial metastases from MRI

As the management of multiple intracranial metastases is rapidly evolving, the demands for detection and delineation of a large number of potentially very small metastatic lesions in the brain on 3D magnetic resonance images (MRI) are increasing dramatically. Artificial intelligence (AI) systems can assist both the radiologist as well as radiation oncologist in their roles in management of patients with multiple tumors metastatic to the brain. In response to PAR-20- 155, we have assembled an academic-industrial partnership including investigators from three academic institutes and an industrial AI team to develop, translate and validate AI systems to address this unmet clinical question. In this proposal, a neural network system based upon multiple scale 3D fully convolutional one-stage objective detectors containing segmentation heads will be optimized and investigated. Training and testing data will be provided from clinical images of patients treated with radiosurgery to multiple small metastases acquired from three academic centers, curated by experts, and augmented by addition of realistic synthetic lesions injected into images. The clinical utility of the network will be investigated for its ability to assist a) radiologists in detecting multiple metastases accurately and efficiently, and b) radiation oncologists in delineating multiple metastatic lesions to support selection of therapeutic strategies and planning of treatments.