Precision Tracking of Low-Grade Diffuse Glioma Progression Using a Novel Deformable Image Registration Approach
Abstract
Purpose
Low grade diffuse gliomas are slow-growing tumors that must be monitored longitudinally to detect progression to more aggressive higher-grade disease. However, the slow growth of these tumors can make it difficult for radiologists to appreciate subtle tumor changes on a scan-to-scan basis. Consequently, progression can be missed or diagnosed too late, delaying intervention. Volumetric tumor segmentation has been proposed to monitor glioma growth but lacks longitudinal consistency, which limits real-world applicability. New, precise, and sensitive methods of detecting tumor progression are thus needed to address this gap.
Methods
The proposed deformable image registration approach pairs a multi-channel, group-wise registration method with precisely matched blood vessel bifurcation landmarks. This approach utilizes all suitable images from a given patient, adding important temporal context, alongside verifiable anatomy. We developed an nnUNet segmentation model trained on biologically realistic synthetic data to segment blood vessels in real patient scans. Vessel bifurcation landmarks in these scans were matched using direct affine registration of the vessel masks and refined with a multi-stream pseudo-twin deep learning model. These bifurcation landmarks are weighted in the registration loss function alongside image information from multiple contrast types to describe deformation on a scan-to-scan basis. The landmark matching method was applied to 5 low-grade diffuse glioma patients as a proof of concept, each with 5 different imaging sessions.
Results
On average, 159 matching vessel bifurcation landmarks were detected between the different image phases for each patient. Phantom studies estimate the landmark pair precision after refinement to be <1mm.
Conclusion
The sub-millimeter precision of the bifurcation landmarks will permit precise detection of volumetric change indicative of progression to high-grade disease while avoiding the more significant uncertainty involved with segmentation. The approach can serve as a powerful tool for radiologists to both intervene sooner and improve patient prognosis.