GLIDE-Reg: Global-to-Local Deformable Registration Using Co-Optimized Foundation and Handcrafted Features

A research team from UCLA and other institutions has published a new paper on arXiv detailing GLIDE-Reg, a novel method for deformable medical image registration. This process is critical for aligning sequential scans of the same patient over time, used in applications like tracking tumor growth or evaluating treatment response. The core innovation of GLIDE-Reg is its hybrid approach, which jointly optimizes a registration field with a learnable module that compresses embeddings from a Vision Foundation Model (VFM). These global, semantic cues are then fused with traditional, handcrafted local descriptors called MIND features. This fusion aims to overcome the limitations of current methods that often lack robustness across different spatial resolutions and anatomical coverages.

The technical results are compelling. On two public lung CT datasets (Lung250M and NLST) and one institutional cohort (UCLA5DCT), GLIDE-Reg achieved average Dice Similarity Coefficients (DSC) of 0.859, 0.862, and 0.901 across six anatomical structures. This represents a relative improvement of up to 3.0% over the state-of-the-art method DEEDS. For Target Registration Error (TRE), a key metric for landmark alignment, GLIDE-Reg scored 1.58 mm on Lung250M landmarks and matched DEEDS at 1.11 mm for nodule centers in the NLST dataset. The strong performance on nodule centers is particularly significant, as it demonstrates the model's robustness for the challenging downstream task of lung nodule tracking—a vital prior step for early-stage lung cancer diagnosis and monitoring.