Abstract
OBJECTIVE
To develop a fully automated segmentation application that provides comprehensive volumetric measures of optic pathway gliomas secondary to neurofibromatosis type 1 (NF1-OPG) and can accommodate diverse MRI acquisition protocols and platforms from multi-center clinical trials.
METHODS
High-resolution volumetric T1-weighted brain MRI sequences, acquired as part of routine clinical care in children with NF1-OPG, were gathered from three institutions, each using a different MRI acquisition manufacturer’s platform: Children’s National Health System (CNHS; General Electric), Children Hospital of Philadelphia (CHOP; Siemens), and Children’s Hospital Colorado (CHC; Philips). Volumetric measures of anterior visual pathway (AVP; optic nerves, chiasm, and tracts) were calculated using our novel fully automated segmentation and analysis method. This method utilizes a deep learning convolutional neural network architecture that captures image context and localizes the AVP in MRI of the brain. Once trained, the automated pipeline takes less than 10 seconds for the complete quantification of AVP on a GPU-enabled computer. The average overlap between manual and automated segmentation (DICE score) as well as the volume error was calculated using eight-fold validation criteria across MRI platforms.
RESULTS
Thirty-two datasets met the inclusion criteria (CNHS=13, CHOP=13, CHC=6). The average voxel size differed across sites: CNHS (0.47 × 0.47 × 0.5 mm), CHOP (0.82 × 0.82 × 0.9 mm), and CHC (0.34 × 0.34 × 0.5 mm). The preliminary DICE score was 0.658 ± 0.052 with a volume error of 0.864 ± 0.476 ml (9.44 ± 8.19%).
CONCLUSION
Using deep learning technology, our fully automated segmentation application has potential to provide volumetric measures of NF1-OPGs that can accommodate cross-platform MRI acquisition protocols used in clinical trials.